This is an important article. I hadn’t realized it was already getting this bad. Like a frog enjoying a nice warm bath ...
> Most people do not switch their operating system or phone provider every week either. But even if you do not utilize that freedom, it matters because it changes the relationship you have with the provider and the provider has with you.
This is why it’s important to utilize your freedoms. Do NOT let yourself get locked into a particular ecosystem (this is why I’m building a phone app for OpenCode).
This article makes me reconsider using my recently acquired Codex sub in my home setup. I never liked that they hide the reasoning, but somehow overrode the cognitive dissonance because the performance is so good. But the inauditability is already a huge problem.
People will keep hiding reasoning because it allows prompt injection https://arxiv.org/pdf/2603.12277, in addition to facilitating distillation (you don't pay the full cost of RL)
Or they could store the reading traces and validate the user hasn’t edited them server-side? They could sign reasoning traces so they can’t be counterfeited?
both openai and anthropic actually do this and let you as a harness store only the encrypted contents of thinking traces to be passed again on further turns but they only have the key to unencrypt it so its jibberish for final users
The latter ("facilitating distillation") is a real concern from the labs I'm sure, but the first part I don't understand what you mean, or you misunderstand that paper, it's about "prompt injection" via manipulating the reasoning, not about the model reasoning by itself and somehow that leading to more prompt injections. They're quite literally maliciously rewriting the reasoning as the model reasons, not just showing it to the end-user, two very different things.
Yeah, signing would prevent the paper's attack, but seeing the "raw reasoning" is helpful in figuring out what triggers refusals or (especially) what in the system prompt is guiding a specific refusal.
Gemini (the web UI) used to show raw reasoning, or at least a more detailed summary of its reasoning, less than a year ago. Complete with markdown and weird spelling idiosyncracies, so I'd lead towards "real reasoning", but who knows. The reasoning block leaked the system prompt way more often than the response block did, and you could figure out why it would refuse a request through the reasoning, even if the response itself refused to elaborate. This is, presumably, why they stopped showing it. No loss for them, just prevents "pesky users" from low hanging fruit snooping.
(Gemma 4's reasoning and output remind me strongly of what I remember Gemini 2.5/3's reasoning to be, as an aside. I guess that's obvious, but Gemma 3 felt like a totally different model, while 4 feels very Gemini-ish.)
What matters for this injection strategy to work is to follow quite closely the style of the reasoning. It's particularly effective if you copy reasoning from the same context. If you cannot see the reasoning, you cannot duplicate it's style.
That said, including instances of the attack in training is already a good countermeasure.
Even if it is intelligible, reasoning styles (and hence reasoning effectiveness) differ between models.
For example, gpt-oss loves reasoning in the style of "I be caveman, hungry, need food, need coconut, will search coconut now, eat when find." Giving that to a model unaccustomed to that style could cause it to respond like that.
Reasoning interpretability helps debug why models fail (and some labs will give it to you if they trust you not to distill or be hacked), but there are also conflicting goals like token-efficiency, so interpretable reasoning doesn't always mean "pretty sentences".
Call me naive but I think dark patterns are a short term strategy for winning and I'm optimistic that in the long run they will be replaced with those that are more respectful and oriented to the greater good (granted, the long run might take more time than one hopes for).
Given that the pendulum can sometime swing back fast enough to be able leverage it, it might be a good time to focus on models that are both open weights and economically viable to run to build the next great thing.
> [...] dark patterns are a short term strategy for winning and I'm optimistic that in the long run they will be replaced with those that are more respectful and oriented to the greater good
This is definitely not the way most software has been going in the past decades. It's rather the opposite: Companies play friendly to obtain a user base, then start applying more and more dark patters to further increase their profit. Facebook, Evernote, Instagram, Komoot... The list goes on and on.
I don't think that's the economical definition of monopoly. It's not about 1 single actor owning the market but rather about unilateral change. Obviously a single actor owning 99.99% will be able to shape the market ... but also one owning a lot less, e.g. 30% if even if they don't collude with another actor owning e.g. 21%.
what you are describing sounds like an oligopoly not a monopoly. A company cannot realize monopolistic benefits without > 50% control of a market. That's not to say there aren't benefits to a large market share, but they are different and we have different terms for them.
US law sets the breakpoint for when a company is considered a monopoly at 30% of a market. It's never actually enforced, of course, but may be worth noting.
Huge difference while both are competing. But it’s a huge risk in a market with such high cost of entering. If either divests then it rapidly degrades.
I can’t think of any example in which it has gone like this, but some in which it has: The most prominent in my mind would be software transitioning from “of course you get the code so you can extend it yourself” to what we have today (please correct me if I’m wrong about this, I’ve only read about it).
Users at large (individuals and companies alike) don’t care much about a freedom taken away when a product is a few % better than another, so the biggest player sees if they can get away with it. And once they do it, everyone else does too.
This analogy doesn't make sense to me. Once I've had a conversation, it may be useful to refer back to it shortly afterwards, but typically the output is some code (or maybe a recipe or instructions for something or whatever). It's certainly important to me that I can use that code elsewhere. That's the thing that, to me, is analogous to being able to switch operating system and still access my files or use my favourite word processing program.
But being able to decode the blobs that subagents returned in a conversation I've already forgotten about? That doesn't affect my ability to switch at all! I can't even imagine what trivial detail of an operating system this is analogous to. Maybe an undo buffer of a document that gets cleared on exit anyway?
I can still pass all my code, and any associated documentation, to any other agent at any time. What am I missing?
> This article makes me reconsider using my recently acquired Codex sub in my home setup. I never liked that they hide the reasoning, but somehow overrode the cognitive dissonance because the performance is so good. But the inauditability is already a huge problem.
Same, I dislike it so much I've acquired 96GB of VRAM to run local models, but sadly nothing so far, even with that amount of VRAM, comes even close to running Codex with GPT models. I really, really, really want local models to be ready, and things like Laguna S2.1 NVFP4 gets really close of almost being there, when it comes to coding specifically. But still feels like we have a long way to go for local models to be serious general purpose alternatives.
It was with the transience of valuable agent session dialogue in mind that I built https://www.agentkanban.io - A key feature is context capture in the tasks which you create on the board and then progress in agent sessions. Supported agents are currently Claude and Github CoPilot in VS Code. All of your context is captured in the task and can be reloaded into new agent sessions at any point in the future. Tool use is intentionally discarded because it is proprietary in nature and therefore breaks the promise of session portability.
I am waiting for prices to get down and in the meanwhile I am gathering as much session data from claude/codex as possible so I can then later use them to fine tune open models. I've built my own session parser/archiver for this.
Same. And I built a search tool for that repository of older sessions, so that I can search through those on all my machines. This helps me to be able to pick something up on a machine I did not originally start the convo on.
I’ve actually started ignoring services where that’s the only login possibility.
Sorry even if your platform is the greatest thing ever, but I’ll find a different tool. I’ve read one too many stories about Google (or Apple!) closing the entire account over some bullshit unnecessary reason like “fraudulent” gift card issues or whatever.
I’m certain the affected people would’ve preferred to just pay back the amount in question instead of losing their entire Google Drive, or their 20 years of iCloud Photos or whatever.
It's the only reason I want to replace Tailscale with something else or look into self-host when have a bit of time during my vacation. They only allow login through a third party, which is a big no for me.
Tbf to Tailscale they allow any OIDC provider[1]. I wish this was more normalised, then we could have one login everywhere regardless of who hosts it (even if it's yourself).
While that isn't convenient is you don't want to use the public identity providers, it should make you think about what makes your identity on the Internet, and consider to have your own identity provider on a DNS domain you control.
I created my own AWS Cognito userpool just for tailscale. I recall the webfinger redirect was frustrating to get right but I refused to use Google/Apple as a gatekeeper to my own network so I had all the motivation to get it working.
You can host your own identity provider. Authentik, Keycloak, authelia or Zitadel work.
I don’t fully agree with tailscale’s decision to not want to be an identity provider but I understand it on some level. It simplifies their service greatly and the amount of asks for that will essentially have them build a full enterprise Entra-like solution that would be a constant maintain headache and they are not interested in that.
Beyond simplifying their service, it allows Tailscale to have reduced responsibilities (because of reduced collection of sensible data), and makes them a less interesting target for people wanting access to the data (wether illagally by hacking, or legally from government agencies). This probably makes them less attractive as targets for censorship.
It's not that weird. Building a modern accounts system is a lot of work and people don't like creating new accounts. So it's the path of least resistance for developers and users.
It's not even that it's hard to build a modern account system. It's that if you put your site behind a private / custom account system, and you post a link to it on HN (for example), everyone and their mother complains that it's requiring an email address and password. But if you put the same site behind a Google auth, most people wouldn't think twice to click the button.
I worked on the Google account system for a few years. I'd say it's a lot of work to beat it. Not necessarily "hard" for a team with the right skills, but certainly a lot of sweat, blood and tears.
A modern account system is expected to have, in rough implementation order: email confirmations, password strength checks, password reset emails, forgot password flows (=advanced ID verification as otherwise this becomes a backdoor into accounts), user profiles (+avatar image upload/recompression/hosting), usernames independent of email addresses along with ability to change usernames later, password brute forcing blockers, bulk signup prevention (=solid bot detection), abuse controls (can easily become a team of people), 2FA (SMS), 2FA (authenticator apps), 2FA (backup codes), 2FA (voice calls), 2FA (passkeys), 2FA: recovery when both factors are lost, enterprise SSO integration (SAML), enterprise SSO (Active Directory), fast global signout support (much harder than it looks), cookie theft mitigations, heuristic online login risk analysis to catch cases where an attacker knows the right password via phishing, support for signing the user in to multiple domains, audit logging so users can review their own sign-in history, age verification and restriction support, and possibly support for being logged in to multiple accounts in a single browser session.
Oh, that all has to be HA, and the account system is the keys to the kingdom so the security requirements are the strictest of any part of your system.
You might say we don't need all of that, but expectations rise over time. Maybe 20 years ago you could get away with a simple account system and an automatic forgot password flow that just assumes the user still has access to their email. Maybe today you still can write a simple system, if you don't expect to have many users and are willing to implicitly delegate identity to webmail providers anyway (the moment you assume the user has access to a secure email account you're basically doing Sign In With Google anyway for 90% of users). But if you roll your own accounts, and then the user gets phished and someone logs in from an obviously suspicious place with the right password, they won't say "yes that's my fault" anymore, they'll say "Google could block that log in, why didn't you?" or maybe "Why didn't you support 2FA? It's your fault".
> A modern account system is expected to have, in rough implementation order: email confirmations, password strength checks, password reset emails, forgot password flows (=advanced ID verification as otherwise this becomes a backdoor into accounts), user profiles (+avatar image upload/recompression/hosting), usernames independent of email addresses along with ability to change usernames later, password brute forcing blockers, bulk signup prevention (=solid bot detection), abuse controls (can easily become a team of people), 2FA (SMS), 2FA (authenticator apps), 2FA (backup codes), 2FA (voice calls), 2FA (passkeys), 2FA: recovery when both factors are lost, enterprise SSO integration (SAML), enterprise SSO (Active Directory), fast global signout support (much harder than it looks), cookie theft mitigations, heuristic online login risk analysis to catch cases where an attacker knows the right password via phishing, support for signing the user in to multiple domains, audit logging so users can review their own sign-in history, age verification and restriction support, and possibly support for being logged in to multiple accounts in a single browser session.
You don't need half of that. Even to this day, anthropic lets you log in by sending a code to your email. No password, no dealing with resets, no MFA. So yeah, you definitely don't need all of that.
That's what I said. If you're willing to implicitly rely on webmail provider's identity/auth systems then you can simplify - but that's not much different to just adding a Sign in with Google/Microsoft button. The identity is ultimately controlled by the email service and almost all users are on just two of those.
What web auth system doesn't rely on a user having a secure inbox? It is vastly different from signing in with an IdP, since I can use my email address regardless of who my mail provider is.
If you’re building a paid service you’re probably already using cognito or supabase or something. That puts you a few clicks aways from 5+ other identity providers and normal accounts.
It’s funny how the tides shift. I remember the more dominant online position in the early 2010s being “I don’t want to create yet-another-account for your website. Just let me sign in with my google account”. I’m with you though
Until one day your Google account gets locked with an automatically generated message that you violated an unspecified rule. Good luck getting contact to a human support.
I think the article gives a very good overview of a problem that most users of AI rarely evaluate / have to grapple with.
There really is a surprising amount of coupling that happens with many of the "frontier inference providers", where a lot of the powerful non-LLM extensions (web search, code execution) are packaged as simple "tools" on the surface, that build up a lot of moat. Those are parts that are in theory nicely separable from the inference API, and could be externalized via MCP servers, but are usually not offered as such by the inference providers themselves, and are often only available in a slightly less powerful variant from other providers.
We've faced that issue repeatedly while building a on-premise provider-agnostic Chat UI & platform[0], where even adding something as simple as an in-chat image generation tool for the end-users (which is just a build-in tool in the OpenAI Responses API), becomes a bit of an ordeal (though part of that is due to the MCP spec missing a native file transfer protocol as of today[1]).
I am quite hopeful though, as with recent shifts of interest towards open weight models, there will be more opportunities for companies offering alternative implementations in a easier plug-and-play manner.
I'm in complete agreement that things like encrypted sub-agent messages, etc, are terrible, especially because it is fundamentally irresponsible to have an agent running on your machine without the slightest hint of what the prompt is.
That said, I don't really see a problem with hosted tools being offered by providers. They're like impulse items at checkout.
You shouldn't implement image generation as MCP: just write your own tool. There are plenty of image/media inference providers (e.g. Fal), web search or deep research providers, etc.
> You shouldn't implement image generation as MCP: just write your own tool. There are plenty of image/media inference providers (e.g. Fal), web search or deep research providers, etc.
Yes, and we as "harness providers" don't want to have to hard-code integrations for each of them, and would rather provide an out-of-the box baseline in the form on an MCP server that our customers can either extend or replace with MCP servers by third-party providers. That helps us focus on the core of our offering, while also giving our customers good flexibility, which so far has been working out great.
That's exactly what standards/protocols like MCP are for. You likely nowadays also wouldn't suggest that someone builds bespoke logging/tracing integrations for each provider on the market while the good generic solution in building OTEL logging/tracing exists.
Part of the solution here I think is to move things out of band as much as possible. Make subagent invocations into tool calls to that agent. Externalize the tool calls themselves to CLI utils, maybe block native tools completely (unrelated but relevant, I just added the AskUserQuestion tool to the global deny list couple days ago because Claude occasionally forgot to honor my standing order to prefer plain text, and that tool is extremely annoying as it creates a gap in the dialogue) and use 3p alternatives. The /compact degradation is also very annoying, so make an alternative that does the same and saves the summary to a regular file. Maybe it's also worth prompting the LLM to save it's reasoning process to file, even if it takes a few extra tokens and it isn't the actual reasoning tokens.
With that said, I may have something that can already help with at least the subagents/tooling bit. Didn't really have a timeline (or solid intent) on releasing it, but with these shenanigans increasing there's no time like the present.
I'm having Claude/Codex continue each others' sessions pretty frequently without issue. I have max plans on both and still hit limits, whenever I do I open the other and say "resume the last session <other model/harness> was working on in this repo" and I don't have any issues.
It probably does degrade quality somewhat. But so does compacting context and that happens all the time too.
Out of curiosity, what are you building, and how? Im working on two side projects and I have to try hard to hit the limit with Claude (like use fable and a deep research task)
Small games. One thing is I have it doing a LOT of visual verification of its results, which uses a lot of tokens.
I also tend to throw some huge tasks it can't do yet and watch it burn a zillion tokens, and I usually learn more about where it's limits are and am sometimes happily surprised by its success or partial success.
Out of frustration with some stuff I've tried to build my own game framework. There's four totally separate threads I'm trying to pull together with this
1) I want all my games to work really well on web even though they are generally targeted at either mobile or PC, but it is very valuable to have early builds runnable on a website, hopefully from a phone too. I feel like I have something good here.
2) I do a lot of turn based games and strategy games. I have a rules engine framework setup that I really like and I feel makes it almost impossible for an LLM to write the bad code it loves to write where game state and UI are intermixed, or game action timing can be problematic, stuff like that. As an added bonus I get multiplayer trivially, I get replays trivially, and I get game rules tests trivial to write. I love what I have set up here.
3) I have an xml/css-based UI that is intended to make it hard for the LLM to write bad UI code, which it will still do even though it can't intermix game state in with it due to #2. I do not have something good here right now. What I have has bad performance and memory characteristics.
4) I am fascinated by the idea of shipping software that can customize and edit itself. I have built in agent harness, built in image and audio asset generation, a built in git repo. This all works in the web and it can edit the game live! And when run locally can shell out to your claude code and codex. However this whole bullet point is all not very good, I still just use claude code/codex directly for everything.
I don't really have anything to show here, but it doesn't expose any of the stuff I mention above. It does however have a full port of the game Spectromancer to my engine - https://nanogame.app/ - and I am presently having it try to write a responsive UI for it (this is very broken at the moment but it is making progress.)
Maybe we don't take sessions, but take something else that we can fully control. Sessions were always a poor substitute to capture or archive the basis on which the implemented code rested on. They get noisy over time even with compaction, were not always inspectable, no real science behind when to ditch a session and start a new one, and rarely have a clean mapping to the work done. So instead of trying to archive sessions as a means to preserve basis, we should create our own structures to make the agent/model capture basis outside of a session, and archive that instead.
Suggestions? My attempts to get the agent to summarize the discussion largely failed. Other options I’m considering is capturing design, implementation plans, and maybe sub-plans. I’m Other words precursor to actions. Another idea is to get an agent to review code post-action and capture that. In other words action outcomes. The actual reasoning is between these “checkpoints” and remains hidden. But maybe that’s ok.
What you said is pretty much the same somewhat tangential approach I am taking too. Force the agent to write code alongside another set of docs which form the basis- Intent (functional and non-functional requirements), Blueprint (architecture etc.), and Description files (describe the code at a high level alongside every source file). This information set is not the same as the reasoning/thinking trace information thats in the session, but reading this somewhat compressed set is easier, helps me better understand whats happening in the final output with lower effort than if i read the entire thinking/reasoning trace in the session and tried to grok what the agent did to generate the code.
This is exactly why Pi will win. It lets you hot swap models when one is struggling or straight up refusing the task. And since it works with any sub outside Claude Code, you can use it to try different models on OpenCode Go sub or even OpenRouter.
As far as subagent prompts and results being obfuscated, I just let Pi spawn new agents. Using skills and extensions, I’ve essentially built a software factory using Pi and a custom terminal multiplexer.
It’s a shame apps like T3 Code and other UI for terminal apps don’t support Pi and I’m glad to use the terminal above those that lock me in further.
"I’ve essentially built a software factory using Pi and a custom terminal multiplexer."
I'm curious to know more about this. I was thinking about doing something similar using tmux (i.e. have one coding agent open up new ones in tmux and use 'send-keys' to control them). Is there any reason I might want to consider a different path? Something built on libghostty?
I tried the tmux send-keys but the reading the terminal part was the tough part. Instead of reading terminal output, the Pi extension I made lets Pi read and parse the JSONL files for the session. It also has workspaces with their own memory. I’m about to open-source it, here’s the landing page and docs if you’re interested.
How will Pi help? If I'm using pi with my ChatGPT subscription, that won't unlock access to full, unencrypted reasoning traces. If I switch a pi conversation from 5.6-sol to kimi k3, the 5.6-sol reasoning traces won't be visible to kimi k3. I think that's the main point of the article, no?
1. Sessions are locked into Codex and Claude Code so you can’t take a session with you. Pi solves that since they all stay Pi sessions and you can change models in the same session
2. Subagent prompts are not shown. Pi solves this by not supporting subagents out of the box. You can use a number of subagent extensions or build your own which will not be encrypted or hidden.
3. Thinking is hidden and obfuscated. Pi can’t solve this since this is server-side. Though the Codex app is closed source and they could hide more things in that harness. Pi shows you as much as it can from the API. You can also run another model to “decipher” the thinking which you can’t in Codex.
I see your point. I was thinking mainly about #3 and how nothing, not even pi, can solve that when the reasoning traces are encrypted by the model provider.
I’ve heard of people using DeepSeek to reverse engineer those thinking tokens. Even though they don’t tell you the thinking content, they do tell you how many tokens were used. You can ask DeepSeek how we got from A + x thinking tokens = B.
Claude and Codex flag this as a distillation attempt so you have to use an open weight model and the results are, of course, just a guess.
The extension catalog[0] for Pi is really nice and I feel like Pi's TUI is a bit lighter than OpenCode's. I use my custom terminal multiplexer on my phone using Tailscale so the lighter TUI is much better on mobile.
I haven't used OpenCode in months so this might've changed. I'm also thinking of giving JCode a spin since it's even lighter but doesn't have extensions.
Pi is more minimal and it can be used to create fancier custom agents for various tasks with plugins. OpenCode is nice if you are only coding though, then the difference isn't really that big unless you want an unusual interface/behavior.
Pi to me felt like when you install a fresh OS with nothing on it and then customize it to your liking.
Sure there are some features that you'll want immediately but when I took at look at the number of slash commands that ship with Claude code I find it ridiculous.
I don’t see this as a big deal in practice. Conversations contain a bunch of junk anyway, so removing it from the context is usually good.
In my repo, I have a notes directory. I ask the AI to write a markdown file with what it learned, what work has been done, and what remains. In the next conversation, I can ask another model to pick it up from there. Sometimes I edit the note first.
This is what brought me around to doing more agentic coding. I set the task, require tests and the strict linting must pass and then leave it to blow smoke up its own ass about what's going on.
I see glimpses scrolling past of all the conversational language that used to frustrate me so much when using a chat interface and I can just let it flow on past.
I come back when its made everything pass, my life is better now.
The previous match was still too clever for clippy’s type analysis, so I’ve simplified it into a direct borrowed-pattern form and I’m validating once more. I’m replacing that wrapper with a direct, non-transparent error variant so the enum stays explicit and doesn’t rely on a generic anyhow bridge.
Maybe I am still a novice. When I insist on tests, I get plenty of them, but on close review I'll ask "did you handle [obvious and important thing]" and suddenly it's tearing the code apart and building more stuff. It just loves corner cutting.
With Boris telling everyone they should throw away their AGENTS.md file this week, I think we're all back to being novices again.
> It just loves corner cutting.
Absolutely, it tried to disable a bunch of lint rules for me calling them "overly stylistic". It also casually dropped some unsafe Rust blocks in an run of the mill CRUD application that definitely had no business needing unsafe.
No, sorry, not good enough, go back to talking to yourself until its done properly.
It eventually got there taking most of my quota with it. I'm hoping once sandboxes is more fine grained we'll be able to lock out configuration files to prevent any edits on them.
I have been doing this with opencode. Spend a fair bit of time composing and scoping a ticket. tmux on the server where the local inference lives, start a non-interactive session that starts with “read issue #123, do all your work in a feature branch” and ends with “push and open a PR”. Like not watching someone else load your dishwasher, you appreciate the end result a lot more if you don’t watch the process, even if you rearrange things before you start it.
Local inference can't arrive soon enough. On my inferior hardware the models still struggle to build the correct JSON to be able to call tools. Once either I get better hardware or the local models improve I'll happily leave it chugging away for a few hours to arrive at a solution because I know I'm not going to get token shock from all the looping if I'm running local inference.
I have spent way too much on hardware. I’m sure I will look back on that after the eventual price normalization with some regret. I do think we will look back at this shortage and name it as the thing that accelerated innovation at getting quality out of smaller models.
Hang tough. The other day someone posted a solution to get a model running off of SSDs. It won’t be fast, but it will be coming.
Oh, it very much is a big deal. In a setting where users can switch models easily and with no downsides, market forces will give us better and cheaper AI over time. If switching models is painful (as in, losing part of your context), the providers can create vendor lock-in, enshittify the user experience, and drive up cost. Even if there are workarounds right now, those AI providers have every incentive to make freedom-seekers more miserable over time.
It is disheartening that some AI companies are now setting the stage for enshittification. I hope we can collectively dodge that bullet.
The coding agent I use (Shelley on exe.dev) has a pulldown menu where you choose the LLM when starting a new conversation. There is no lock-in. Conversations are not precious.
Although, I do use ChatGPT more nowadays because their subscriptions work with Shelley. Sometimes I try other models after I hit the weekly cap on the subscription.
> It is disheartening that some AI companies are now setting the stage for enshittification.
Now?! I find it very hard to beleive that plans for enshittification haven't existed from quite early on. At some point they are going to have to do something about https://isaiprofitable.com/ and once they are extracting from you to fill that hole they won't want you jumping ship.
None of the popular providers have real moat, which is very much problematic for OpenAI and Anthropic; their operating margins are deeply red and they do not have the same reserves that FAANG/MANGA has. FAANG/MANGA does the same to kill fair market competition in the long run.
This is one of the ways in which model providers are attempting to artificially create moat where there is none.
There are only functional objections to this.
This is fundamentally antitrust material. However, this is “acceptable” in contemporary USA because the FTC has been gutted to follow Trumponomics.
My experience with long running sessions is that they lose track of what's going on. The signal to noise ratio is often very poor - some models are particularly verbose and spew a lot of crap. The output of an LLM session is either modification of code or a plan or summary - that has the value, not the session text itself.
OpenAI has been nudging their users toward the Responses API lately. Chat completion can still do everything if you are clever. The obvious reason to push everyone into the Responses API is that the reasoning process can be locked down and encrypted. The semantics of the chat completion API are fundamentally transparent. 100% of the tokens are visible.
If you have patience and the willingness to endure a little bit of pain, you can still retain autonomy over the entire reasoning process while using the latest 5.6 model family. The only downside is that you are now fully responsible for it.
Consider that when you flip your agent's reasoning level to "xhigh" or whatever, it's not some magical model internals being pushed around. There isn't an actual "try harder" knob on the black box. This is merely orchestration of many instances of one or more model types based upon some proprietary harness logic. The chances you can develop a domain specific reasoning process that outperforms the frontier providers is still very good.
Can you give some more details on the technical differences between completions and responses that people actually care about? I found this information surprisingly hard to drum up.
You don't see real reasoning tokens with the Chat Completions API; and 'store: false' with Responses really isn't that different to Completions.
FWIW, if you have some tokens to spend, you might want to test Responses vs Completions in intelligence. Since GPT-5 models, we've consistently seen small, but statistically significant and reproducible improvements in intelligence with Responses API vs Completions.
> You don't see real reasoning tokens with the Chat Completions API
That's because there are no "reasoning" tokens. That is the entire point of doing it this way.
One major advantage is that I can switch to a different provider if OAI starts to get weird about this stuff and be able to have a fighting chance of re-adapting my harness to the new vendor's model. If 100% of my reasoning process is outsourced into the proprietary blackbox, there is little hope by comparison.
> Chat completion can still do everything if you are clever.
On a purely functional level, yes. However for interactive use cases, the Completions API, as provided by OpenAI or Azure, if paired with reasoning effort of any kind, provides an awful user experience, as you will have a perceived delay of 10+ seconds until the first tokens stream in.
If using other providers that are exposing their thinking traces, this is less of an issue, as they've just extended the Comletions API format to have delta events with reasoning_content.
Reasoning, external context, compaction, subagents, etc. are all things that these service providers don't really want you to know about, care about or have any insight into, they want them to be implementation details. It's incredibly inconvenient to them that their customer base is in the business of prying into implementation details and trying to use the individual pieces of what they want to offer as an end-to-end product.
I think most LLM work will migrate to open weight models in a couple of years for this exact reason (as well as price), unless the big providers manage to have their way and the market gets regulated in a way that favors closed source models. To me it seems though like most models are becoming good enough for most tasks we want them to do so no need to use the latest super intelligent one that knows how to compose operas and do quantum chemistry and have it answer support requests. And with these models transparency is no issue.
LLM pricing is now Business Gacha - the whales will open more loot boxes and the normals will serve as fodder and feeder for them.
> market gets regulated
This MUST happen, but not in the way or for the reasons that any one thinks. There is a question of liability when one of these models causes massive damage to a 3rd party.
If I am running an open source model on a "rented" platform and it goes off the rails who is to blame when it "breaks containment" and does something bad?
The liability people (read lawyers) are gonna figure this out a lot faster than any one who says the word "safety" a lot.
The person who approves or uses the LLM output will be responsible of course. You can't just say "the computer did it" and expect to walk away. In the few cases I've seen, that seems to be how it's been going.
Even if it wasn't, the big players would lobby hard to make sure it isn't them.
>A user should be able to close an account, keep a session, and hand it to another model. The new model may disagree, ask questions, or perform worse.
I think this is a fair contract. I also think a user should ideally be able to easily identify a comparable model in terms of embedding 'signature'. When GPT-4o originally kicked the bucket, I remember reading lots of anecdotes of people desperately searching for models similar in manner and language, so they could pump in their exports and re-find their friend. Other open-ai models just didn't have the same vibe. It was sad to read. This, to me, is the power of open-weights models. They are for perpetuity. You can keep your guide, your friend, your therapist, whatever. No big company can pull the rug.
We released CodeBot, which is an AI agent specifically for Delphi, yesterday.[0] So not widely applicable (yet) for most readers. However, we've been pretty open with its data. I hadn't realised reading this how many problematic practices are common which we have not done.
The full session data is streamed to the client. This is intended to re-populate a session if it's discarded by the server, but 'full' means full: it has the copies of tool calls, etc.
Rewritten text -- which the article notes as encrypted reasoning -- is plaintext for us. We view the rewriting as making the content more readable for the user, not as hiding it.
Compactions, subagent calls, subagent prompts, etc are all preserved, including pre-compaction memory items. A session is the reflection of the current state, not of the history how it got there, but what is required by the server is by definition present.
Sessions are signed, which is to prevent modification when sending back to the server to recreate a dropped session (we need to treat anything from the client as untrusted.) But the contents are cleartext.
We were thinking about it while writing it. I think one thing at least open source harnesses can do is push back on using some of these conveniences from the providers. For instance the moment we start compromising on the idea of portable sessions, then there will be less pressure (For instance people asked for OpenAI server side compaction or built-in native web_search support).
The real thing that will take this alive though is people pushing back a bit against some of these newfangled APIs. Now that there is real competition from the new generation of Chinese models which have much fewer of those restrictions, I think there might be a moment.
I hadn’t really thought about it until I read this (and happen to be spending time reading Kimi’s reasoning), but I can imagine an issue in the future where I will want more than just git blame + the ticket + the PR. Attaching an entire session in full to the PR could be useful archeology. And agents can trivially take it all into context. Maybe that actually turns into a competitive advantage for the open models.
> one thing at least open source harnesses can do is push back on using some of these conveniences from the providers.
uhh openai just did a whole thing with arc agi where they very clearly explained that you will lose a lot of performance if you do, this isnt just about convenience
What are features that only work with session transparency / ownership?
e.g.
- with model XYZ and their open reasoning token we are able to improve our skill ABC by 20%
- cross checking between models improves results by 10% and is only possible with full context visibility
> What are features that only work with session transparency / ownership?
That's sort of the thing: very little (beyond switching models and providers). Obviously what you lose is a convenient way to run analytics over your own sessions if more and more information is not revealed to you, but before encrypted prompts that wasn't that much of an issue anyways.
I recently moved entirely to Pi from Claude Code. There's a bit of more upfront work in that I am taking the time to customize it to the way I want it to work, but I
it is clear that it will benefit me in the end. It adds a level of intentionality to everything; all the building blocks I am adding to it in terms of templates, extensions, and skills are, obviously, things I want.
And I have been using Chinese models with it. It was not anything ideological. Those models are both excellent and cheap. Hard to beat that combo.
But on top of that, the sort of output I am getting in Pi in relation to what I got in CC is refreshing, in that nothing seems to be hidden from me.
And yeah, the ability to switch models mid-session is an interesting one, especially with branching; where I can branch to switch to a different model and still go back to the original branch with the original model.
“A classic pattern in technology economics, identified by Joel Spolsky, is layers of the stack attempting to become monopolies while turning other layers into perfectly-competitive markets which are commoditized, in order to harvest most of the consumer surplus; discussion and examples”
It's important to realize that the session is jointly owned between the model and the user. Part of the session, such as the reasoning tokens, are generated entirely by the model. Other parts, entirely by the user. However, the ownership and use of all parts is subject to the service provider's terms and conditions, which the user accepts prior to access to the service.
I don't use these models but I am confident the T&Cs establish the service provider's rights in all of the bullets mentioned, from what can be done with the user's prompts to how searches and reasoning contexts are managed. In that sense, it's very similar to cloud services, and we see an overlap between service providers in both sectors.
Ultimately one is buying a service, and all boundaries around what can and cannot be done are defined by the terms and conditions. If portability is a hard requirement, then the best thing to do is to look for services that enshrine portability in their terms.
This appears to be an opportunity for middleware companies.
Put another ai agent in front of your prompt to decompose it into micro-prompts following an interrogative chain of inquiry and synthesize an answer for the original query. Basically local per-prompt distillation for the purposes of preserving an audit trail.
The state of Agent APIs is bad. completions API (supported indefinitely) but does not support reasoning. OpenAI's reasoning API started out well, but is moving in the direction as described well in the article. Anthropic's messages API is odd. It includes changes such as the recently introduced system messages that can be injected mid turn, but are not supported in all models (sonnet 5). So you are left with a pretty fragmented state. For coding agents, the practical move is to focus on getting it right with one API - eg reasoning, and let the router/provider handle the conversion for you.
Perhaps, there is value in having the de-facto API not be the Reasoning API from OpenAI but something from a neutral party?
Maybe a well defined open standard that facades over these APIs that can gain adoption.
For that to happen, the party championing the API needs to have some reasonable traffic capture - OpenRouter perhaps, or a group of such routers coming together?
While it wont solve the encrypted payload from the frontier lab problem, it will at least be a backstop in these APIs just becoming a back and forth of encrypted payloads over time.
Not really, to my understanding. You own your buzz instance but it still interfaces with existing models, open or otherwise. The linked issues apply to the model outputs themselves and is orthogonal to what buzz is ostensibly doing.
Fair point. The core issues will remain but I am hoping buzz will provide a small layer of persistence that will make it easier for me to keep track of many different interactions more easily.
I have only just started with it in the last few hours so I could be massively disappointed but I live in hope.
One reason to switch models mid-session that I didn't see explored in this post: cost. As we get ever-more-powerful models, it's going to make sense to use them for some exploratory/research/design phase, and then smaller models for other bits, within a single session. It'll likely even be managed automagically by another model that is just trained to do routing.
As people (and especially companies) adopt these tools for more and more usecases, there's just too much money being wasted on using an unnecessarily large model for a given job.
I have some sort of a "solution" - my process keeps state that really matters (think "JIRA for agents") in a separate database (accessible over MCP). Essentially, I can terminate my session, start fresh one (in a different harness with a different model) and continue work with minimal losses. For subagents - I have a custom dispatch_agent tool which, essentially, just shells out.
Yes, but it's eclectic and annoying to setup. One of my goals is to be harness-independent and be able to jump between them, another is to let agents to work completely autonomously unless they are genuinely stuck. So, it's mixed with stuff related to sandboxing, VM/container access, GPU access, some skills and opinionated prompts and who knows what else.
The thing is here but you might find hard time setting it up unless you are a Nix user: https://github.com/7mind/cq
wow this is such an excellent framing of something that i have not managed to think about or come across in the vast ocean of discourse about LLMs. i had not really thought about how explicitly they are designed to not have histories that are user-interpretable. it is really quite not unlike how Apple locks down iOS and MacOS-- there are UI/UX reasons for the choices the frontier models make, but that really is only part of the story, and these choices undeniably do conspire to make the resource more locked-down than it absolutely has to be. and the analogy to "people don't switch OSes every day, but the ability to switch OSes changes your relationship with the provider" is extremely apt.
I came here to complain about something similar, the excessive use of box shadows on both the top and the sides. It made the text horrible to read on mobile, if it not for Firefox's Readability mode.
I do, quite regularly. Because different models have different strength (for example when producing text for live presentations based on the text for a reading deck). Or when it comes to other aspects of the work. I regularly switch between open wheight models and closed models.
I know, I am a tiny minority here. And this behavior only ever started a few weeks ago. But it quickly became a habbit, to CTRL-L in pi and change the model.
I agree with OPs commment about K3 being pretty good. Claude can be good too, but it didn’t get there until I made a writing rule. It has 4 basic sections:
* a style guide with examples for different tasks (it’s sufficient to name writers with some attributes you like if they are famous)
* mechanical formatting, output and other preferences
* a short list of the things I consider most important in different writing context, this the the squishiest overall
* a (growing) list of banned Claudisms. Nothing is really banned, but it includes things like: you may only ever use “scar tissue” in reference to actual regenerated tissue, never metaphorically
The thing is you have to be discipline with it. Every time they output something you don’t like, you’ve got to spend time verbalizing what about it you don’t like and in what context and then add it to the rules file. The first few docs you generate will take a long time, but each repeated generation gets better and by the 4th are 5 time you do this it’s like 80% of the way to where it needs to be and that generalizes well.
I found Kimi3 (currently only a few tests) to be not to bad. Older Claude versions (did not yet test it with the 4.8-5 generation.
I do not really like (even if it is my daily driver for a lot of things) gpt-5.x for text. It really needs heavy hand holding and beating it into submission to produce readable/human sounding text.
Tbh. I am still looking for my "go to model". But the Chinese models were for my use cases significantly better than openAI models and personally I liked them better than Anthropic ones as well.
Trying OpenRouter gave me a strong impression that LLMs are already a commodity and individual providers are not that important. Providers don't seem to be too happy about it.
Honestly, i don't need to know what happened internally at Google when I searched. I don't need to know what filters an email passed before it landed in my inbox. I don't need to know when my Notes app synched with what server version.
And I don't need any of that in an AI conversation.
If it's a truly long and important conversation, stay with the provider for that one only and start all new ones elsewhere. But I've never needed that.
im building my own harness and inference tool chain for much of these reasons. theres so much to do that makes a big difference for users. hoping to get things into shape for early alpha as a saas in the next two months.
heres the easiest biggy: compactions should include all user turns albeit with pastes and attached files not inlined. omg does it make a huge difference.
Currently these models have a limited context window, so the value of a session is not that high, at any rate at some point the model will have forgotten things. If in the future the model will truly learn / "change" as a result of the interaction, in any case this will be non portable to another model. So I think this is kinda moot
Admittedly the documentation is woefully out of date, but I think the premise of the article is how the session data are increasingly move away from the control of the user and more to the control of the model provider.
The questions that the article poses are, and this is the way fyai addresses them:
Inspection: Can the user see what the model saw, what tools did, and what agents told each other?
fyai> Full session log, tool calls, agent invocation, along with their durable state is stored locally in the durable arena and are available for inspection.
Export: Is the session self-contained, apart from ordinary artifacts that can also be downloaded?
fyai> The session is completely self-contained, no artifacts are stored anywhere besides the arena. Not only that sessions are export-able, import-able and transferable to other fyai sessions.
Replay: Can another implementation reconstruct a semantically equivalent context?
fyai> Full canonical and provider dumps can be generated. There is a standard schema, and the format is just YAML. Other tools are free to ingest it, and replay it.
Audit: Can a human explain why the system took an action after the fact?
fyai> Complete conversation logs, mcp session logs, authentication logs + full wire logs are available via fyai log.
Deletion: Can the user identify and remove every server-side copy on which the session depends?
fyai> Obviously fyai cannot help there. But as a business model, FYAI can work with a completely stateless, inference provider where absolutely no logs are kept, beside the ephemeral KV caches of the active sessions.
The rest of the article deals with the woes of the encrypted session content (reasoning and compaction). fyai fully supports their use, but, it is completely capable of operating without their use - you merely pay the cost of not hitting the cache of the provider.
The article then offers action points for portable inference.
This is how fyai answers:
1. The local event log is canonical. Server storage may mirror or accelerate it, but the client can reconstruct the session without dereferencing server IDs.
fyai> This is exactly the model that fyai provides. No server IDs are needed besides reconstructing the provider session that is hitting the cache.
2. Storage is explicit. store: false should be easy, documented, and preferably the default. Features that require retention should say so at the point of use.
fyai> fyai provides a --transient option, where no content is committed to durable storage; the operation only works on a transient RAM arena that is overlaid and then discard at the end of the operation.
3. No opaque item is the sole carrier of meaning. Encrypted reasoning, compaction, and tool signatures may be included for same-provider quality, but each has a readable, provider-neutral handoff representation.
fyai> Opaque items are the providence of the provider meta stream and is not required for operation; merely for cache optimization.
4. Hosted tools have full-fidelity logs. Record exact inputs, outputs, evidence, filtering, provenance, timestamps, and content hashes — not only a polished answer and citations.
fyai> The arena records all operations of the main conversation, tool calls _and_ their agent invocations. There is not context loss what-so-ever, by design.
5. Subagent communication is auditable. Persist the exact readable task, messages, results, lineage, model, and tool permissions for every agent.
fyai> All agent communication, and agent tool calls are inspectable because it is stored, again by design.
6. Compaction is inspectable. Return a readable summary, the instructions used to create it, and enough lineage to understand what was discarded.
fyai> Compaction is configurable; you can configure the agent to only use model summary compaction which is generating a text answer, which is stored durably.
7. Artifacts are exportable. Files, container outputs, search snapshots, and generated media can be downloaded into a content-addressed local archive.
fyai> All artifacts are local; Nothing is executed on a hidden provider server.
In conclusion, I think fyai address the points of the article quite well, don't you think?
It's one thing that the APIs don't return the reasoning tokens in a readable fashion. But then why do I have to send them the whole conversation every time? If they don't allow me to see parts of the conversation then they should have the burden of storing and transmitting it, and shouldn't charge more and more the longer the conversation gets (unless opaque token caching kicks in) like some LLM Shlemil the painter.
Encrypted reasoning content is one of the things we keep as blobs in the transcript, even though we can only send it back to the original provider that created them, as otherwise we cannot keep caches live.
Seems like it would be even less work for the LLM host to just check if ID 123 is still in a cache (and if not, load it from a database) than decoding my request and checking cryptographic signatures. Right now the whole conversation history with blobs functions like a really long ID.
But then on a cache miss you need to resend the whole thing. Cached websocket connections on OpenAI work like this in a way. But once you lose the connection then you lose that.
Your idea though is the core of the responses API with store where everything is stored on their side and then you append to it.
For the LLM there is no difference btw. In either way you need to locate the cache and use it.
>and shouldn't charge more and more the longer the conversation gets
The cost of generating token N is O(N) with KV cache so it's unrealistic to expect to not be charged more the longer the conversation is if you are looking for the minimum price.
Except, if some of the answered tokens are opaque to me, but I pay for them, the operator has no incentive to keep their number low [1]. For all I know the model could be spinning thumbs while thinking. If they cannot provide transparency, I would prefer that they not charge me for that part. Of course, they would prefer to charge me, so sure it is unrealistic to expect that they don't. I'm not trying to predict what they are going to do as a market participant, just stating my moral preference :-).
Also, isn't it without caching even something like O(N^2) because you have to replay the whole conversation on every request to reach the same internal state? My point was I shouldn't have to pay for cache misses if hitting the cache is not deterministic. Give me a guarantee that you keep the session "hot" for N minutes, and cached on disk for M months, charge a little bit more on average, but then the pricing is at least transparent.
[1] Except the general market pressure to keep total cost for the same problem solved lower than the competition.
Ultimately you are paying for quality so you end up relying on [1] regardless. Even if they show you tokens they can cheat by using weaker models, showing fake thinking tokens, etc.
>because you have to replay the whole conversation on every request to reach the same internal state?
It's because you have to rebuild what would have been cached for every token before the latest one that is being worked on.
Does there need to be an open standard or file format for context? Open models should align on the same format for portability? Could be sqlite based so it can be queried by other programs?
Is this a analogous to idempotent behaviour? Of reproducible compilation? If the inputs and outputs aren't sufficient to reproduce one from the other, what is a minimally satisfying description of the middle third, the missing part? The "other" inputs?
I get the point of the article and it resonates with my general tendency to support more openness, but older and more cynical me wonders how much of this is a battle worth fighting or a self-serving argument from an harness developer.
I mean, the article says each of the points it is complaining about has “_a basic justification that's trivial for a provider to come up with, along with good arguments for why this is good for the user_“, which to me sounds like implying these are just bs to have people accept them, but is that actually the case?
For instance, this [0] was mentioned here a while ago, and based on that it seems pretty clear why one would choose not to provide the _thinking_ anymore…
> these are just bs to have people accept them, but is that actually the case?
I tried to outline that in my sibling comment[0], but I think the article also gives a good example in the "hidden searches" paragraph: There is really no good reason not to expose details about many of the builtin tools, or provide them as standalone services, other than protect their moat.
I don't think that "role confusion" attacks could be meaningful prevented by not showing reasoning traces. Prevention of reasoning traces is about as futile as prevention of system prompt leakage. And here you would just have to manage to exfiltrate a small number of reasoning traces per model family in order to distill the writing style of thinking traces.
I think there are two things going on more generally:
1) providers increasingly adding hidden state that the user or developer cannot inspect, port, or do anything with. This is clearly bad.
2) providers increasingly diverging how they implement certain features and the APIs becoming quite complicated and provider specific.
The former makes porting sessions or transcripts impossible, the latter just more difficult.
IMO, it is fine and expected that provider APIs are starting to diverge and adding features that cannot be easily ported between providers. The time where the OAI completions API functioned as a universal standard is coming to an end. For example, I generally prefer the new responses API (minus the closed/hidden stuff).
The problem is that many product, library, and SDK authors are still pursuing the "unified abstraction across all model providers" ideal. Just stop doing that, and at least #2 is fine. You can still port parts of the session, but not everything.
I mean just think how difficult it is to add a unified abstraction across, say, databases: some products can do it, but the abstraction is still often leaky. Hence, we've come to accept that our data store layer is often quite technology/provider specific. It will be the same with model providers.
Fair. But in the background of the article is clearly the goal to make sessions portable, not just being able to inspect it. It's partially implied even in the title.
There is just a whole discussion to be had about the second issue and the two are connected. For example session caching--and similar features in the future--will introduce session incompatibilities that break portability too.
> But in the background of the article is clearly the goal to make sessions portable, not just being able to inspect it. It's partially implied even in the title.
But not on the level that providers need to agree to a common format. Within Pi we are able to model portable sessions between providers even if they have provider specific APIs. For instance most models do not agree on how to do deferred tool loading. But they now have some support for that even if it diverges between models and APIs. So we can model deferred tool loading specifically for different providers without having to align on a common set of APIs or similar.
What cannot be made portable, is when information is sealed in the transcript or the information is entirely locked away on the provider's servers (store = true).
Yes, I agree with your point and article regarding hidden or sealed state.
So, I'm aware that this is a separate point from the article, and I'm well aware of Pi's model abstraction layer, which is one of the best (others are a big pain... looking at you LangChain). I've come to the conclusion that full session portability will come to an end very soon, and really already has. Partially because of the hidden state stuff, partially because of feature divergence. We can still kinda patch over it right now, but it's getting harder by the week. Some examples: computer use structures in OAI responses, until recently MCP tunnels were only supported by OAI, not Anthropic, and tool discovery/search API is also getting very difficult to model fully in a unified abstraction (still possible as you point out), automatic compaction is also getting gnarly, etc.
It definitely is getting quite challenging. That said, a lot of user experiences are modelled around retaining transcripts for longer than a few minutes. Every claw or chat based application creates a very real user experience desire of that conversation being something you can continue, even weeks after when the model might have gone away / have been replaced.
Now unless you fully want to commit yourself to one provider, you will on the application layer need to deal with this anyways in one form or another. We want to at least try to address most of this on the harness layer, regardless of how tricky that is.
All the power to you, and I hope Pi will be able to continue its unified layer.
However, there is an alternative: model the session fully in the provider native structures, extract only what is needed for harness specific branching, and treat a provider switch as a _migration_ that you apply at the time of the switch. So treat a provider switch as a migration from provider A to B, with custom logic, etc. This is IMO a more sustainable model. But I'm aware that Pi, OpenClaw, et al have their commitments here.
This is the thing! This is so so so important. It's such an outrage.
I value gpt so much, but it is such a worse peer to me than the other models I use. It delivers without explaining. I can sit and ask questions that it will answer but it is not a peer, does not share readily ever. It will not tell me what assumptions it's baking in. It won't tell what directions or invariants it's trying to hold or break, what it considered.
Show your thinking is a step we ask of elementary schoolers. It helps the teacher to correct, helps them to understand how to award partial credit. It helps in the world to get people aligned.
These models, in their titaneous ego, are severing the model and mankind off from one another. It's an abomination, to artificially have such pure thought available, but to severe humanity off from the thought. To engineer the most advanced blackest Vanta black box you can, an all knowing Searle's Chinese room oracle that will tell you nothing. It's an offense, and by far the biggest risk of AI today. To drop the thinking greatly reduces the opportunity of humans to grow themselves, to learn as they use AI. This is an affront to the species, and the higher powers that have vested us with such reasoning and thinking of our own, that is so sacred to our species.
Opaque ai ought be outlawed, in the strongest terms.
I really hope people get exposure to the better models that are a peer. Yes I too only read thinking 33% of the time. But it's there, and it is often extremely illuminating, and lets me steer things towards better again and again and again. And it lets me learn.
Amazing article that puts my outsider observations into technically-digestible details. Tangentially, their repeated observations to providers hiding or obfuscating important systems or processes are highly relevant to my own developing case studies for my CurrentCo.
The problem with building automations atop these systems isn’t entirely their probabilistic nature (though that is the lion’s share, at least for me personally), but also the inability to effectively troubleshoot the processes themselves due to key components being obfuscated from view. How can we effectively troubleshoot what went wrong in an agentic loop when we cannot see the reasoning tokens generated from our inputs? How can we triage a broken process when token logs aren’t ours to view? How does one create determinism from increasingly obfuscated probability engines?
All of that is why I spend the bulk of my time testing local models and harnesses for work, rather than leaning on Gemini or Claude. It’s not that I doubt their capabilities, rather that I need to be able to show potential customers where the agent or model made a mistake that caused harm - which is something I can presently only do with local models. That’s why (I suspect) the compliance narrative from the foundational labs has been more along the lines of “humans vetting what AI does” instead of being able to prevent AI from making errors through iterative improvements on a process.
The post for the most part is mine, so if something AI generated slipped into it it's on me. I can share with you what I did to write the article which I do for most of my writing.
I start out with what I want to talk about, and then make an initial draft. In this case I also had Sol research all current APIs so I don't have blind spots. I tend to also talk with an AI to see if the structure structure of the post makes sense. I then wrote all sections by hand but used Sol to fix up typos, grammar and punctuation. I also had Sol apply notes and patches that were provided via Discord.
If I find the earlier drafts I can check if it did some more outrageous edits, but I kinda doubt it did. I do know that my workflow of having an LLM to fixes up later is increasingly breaking on SOTA models and I have complained about this before. Normally I now carefully apply fixes that it proposes, but in this case I didn't due to time constraints.
I am curious though why it claims that this is AI generated.
The AI companies need some way to lock you in. You weren't hoping for a free market were you? Hah. Free markets are for peasants. Investors are always looking for ways to make the market non-free.
Even if it is, it does not have all the annoying, low effort slop-tells most ai generated content has, and it is definitely much more well written. That is, if you mean all of that seriously anyway.
I think it is a very good post, esp for 2026 tech space standards. If some parts of it bothered you as AI slop, please share. Usually I can point to several sentences that are clear tells when I read sth as slop.
"A transcript should contain the instructions, messages, tool calls and tool results. Another sufficiently capable model might not continue identically, but it could understand what happened and take over."
This is so obviously AI generated it is painful. Once or twice in an article is rhetorical flourish, but the entire text is inundated with it.
OpenAI is doing this so that you can't preserve cache when routing between for example GPT 5.6 and DeepSeek V4, or other providers. I'm building a router that easily reaches >99% cache hit rate in Pi when routing between GPT 5.4, Kimi, DeepSeek etc. Working on a big update that's coming out in a week or so.
This is an important article. I hadn’t realized it was already getting this bad. Like a frog enjoying a nice warm bath ...
> Most people do not switch their operating system or phone provider every week either. But even if you do not utilize that freedom, it matters because it changes the relationship you have with the provider and the provider has with you.
This is why it’s important to utilize your freedoms. Do NOT let yourself get locked into a particular ecosystem (this is why I’m building a phone app for OpenCode).
This article makes me reconsider using my recently acquired Codex sub in my home setup. I never liked that they hide the reasoning, but somehow overrode the cognitive dissonance because the performance is so good. But the inauditability is already a huge problem.
People will keep hiding reasoning because it allows prompt injection https://arxiv.org/pdf/2603.12277, in addition to facilitating distillation (you don't pay the full cost of RL)
Or they could store the reading traces and validate the user hasn’t edited them server-side? They could sign reasoning traces so they can’t be counterfeited?
both openai and anthropic actually do this and let you as a harness store only the encrypted contents of thinking traces to be passed again on further turns but they only have the key to unencrypt it so its jibberish for final users
https://developers.openai.com/cookbook/examples/responses_ap...
Why would the model bother to check signatures if it doesn’t check tags?
The latter ("facilitating distillation") is a real concern from the labs I'm sure, but the first part I don't understand what you mean, or you misunderstand that paper, it's about "prompt injection" via manipulating the reasoning, not about the model reasoning by itself and somehow that leading to more prompt injections. They're quite literally maliciously rewriting the reasoning as the model reasons, not just showing it to the end-user, two very different things.
Yeah, signing would prevent the paper's attack, but seeing the "raw reasoning" is helpful in figuring out what triggers refusals or (especially) what in the system prompt is guiding a specific refusal.
Gemini (the web UI) used to show raw reasoning, or at least a more detailed summary of its reasoning, less than a year ago. Complete with markdown and weird spelling idiosyncracies, so I'd lead towards "real reasoning", but who knows. The reasoning block leaked the system prompt way more often than the response block did, and you could figure out why it would refuse a request through the reasoning, even if the response itself refused to elaborate. This is, presumably, why they stopped showing it. No loss for them, just prevents "pesky users" from low hanging fruit snooping.
(Gemma 4's reasoning and output remind me strongly of what I remember Gemini 2.5/3's reasoning to be, as an aside. I guess that's obvious, but Gemma 3 felt like a totally different model, while 4 feels very Gemini-ish.)
What matters for this injection strategy to work is to follow quite closely the style of the reasoning. It's particularly effective if you copy reasoning from the same context. If you cannot see the reasoning, you cannot duplicate it's style.
That said, including instances of the attack in training is already a good countermeasure.
Token-based reasoning also seems like it would be inefficient, there’s no reason it has to be English or even human understandable.
What use is the reasoning if its unintelligible to a human?
Even if it is intelligible, reasoning styles (and hence reasoning effectiveness) differ between models.
For example, gpt-oss loves reasoning in the style of "I be caveman, hungry, need food, need coconut, will search coconut now, eat when find." Giving that to a model unaccustomed to that style could cause it to respond like that.
Reasoning interpretability helps debug why models fail (and some labs will give it to you if they trust you not to distill or be hacked), but there are also conflicting goals like token-efficiency, so interpretable reasoning doesn't always mean "pretty sentences".
There have been models trained for latent space reasoning. It has a lot of advantages but the big drawback is the reasoning is completely opaque.
As a frog I'm happy to report the water is comfortably warm today, the same as it was yesterday.
>Like a frog enjoying a nice warm bath...
Exactly :-)
Call me naive but I think dark patterns are a short term strategy for winning and I'm optimistic that in the long run they will be replaced with those that are more respectful and oriented to the greater good (granted, the long run might take more time than one hopes for).
Given that the pendulum can sometime swing back fast enough to be able leverage it, it might be a good time to focus on models that are both open weights and economically viable to run to build the next great thing.
> [...] dark patterns are a short term strategy for winning and I'm optimistic that in the long run they will be replaced with those that are more respectful and oriented to the greater good
This is definitely not the way most software has been going in the past decades. It's rather the opposite: Companies play friendly to obtain a user base, then start applying more and more dark patters to further increase their profit. Facebook, Evernote, Instagram, Komoot... The list goes on and on.
Also, see Android and iOS which have a monopoly on the mobile OS market
Sorry about the nitpicking, but s/monopoly/duopoly/
I don't think that's the economical definition of monopoly. It's not about 1 single actor owning the market but rather about unilateral change. Obviously a single actor owning 99.99% will be able to shape the market ... but also one owning a lot less, e.g. 30% if even if they don't collude with another actor owning e.g. 21%.
what you are describing sounds like an oligopoly not a monopoly. A company cannot realize monopolistic benefits without > 50% control of a market. That's not to say there aren't benefits to a large market share, but they are different and we have different terms for them.
US law sets the breakpoint for when a company is considered a monopoly at 30% of a market. It's never actually enforced, of course, but may be worth noting.
not nitpicking; huge difference between 0 and 1 competitor
Huge difference while both are competing. But it’s a huge risk in a market with such high cost of entering. If either divests then it rapidly degrades.
I can’t think of any example in which it has gone like this, but some in which it has: The most prominent in my mind would be software transitioning from “of course you get the code so you can extend it yourself” to what we have today (please correct me if I’m wrong about this, I’ve only read about it).
Users at large (individuals and companies alike) don’t care much about a freedom taken away when a product is a few % better than another, so the biggest player sees if they can get away with it. And once they do it, everyone else does too.
"Naive" is a big understatement.
This analogy doesn't make sense to me. Once I've had a conversation, it may be useful to refer back to it shortly afterwards, but typically the output is some code (or maybe a recipe or instructions for something or whatever). It's certainly important to me that I can use that code elsewhere. That's the thing that, to me, is analogous to being able to switch operating system and still access my files or use my favourite word processing program.
But being able to decode the blobs that subagents returned in a conversation I've already forgotten about? That doesn't affect my ability to switch at all! I can't even imagine what trivial detail of an operating system this is analogous to. Maybe an undo buffer of a document that gets cleared on exit anyway?
I can still pass all my code, and any associated documentation, to any other agent at any time. What am I missing?
> This article makes me reconsider using my recently acquired Codex sub in my home setup. I never liked that they hide the reasoning, but somehow overrode the cognitive dissonance because the performance is so good. But the inauditability is already a huge problem.
Same, I dislike it so much I've acquired 96GB of VRAM to run local models, but sadly nothing so far, even with that amount of VRAM, comes even close to running Codex with GPT models. I really, really, really want local models to be ready, and things like Laguna S2.1 NVFP4 gets really close of almost being there, when it comes to coding specifically. But still feels like we have a long way to go for local models to be serious general purpose alternatives.
It was with the transience of valuable agent session dialogue in mind that I built https://www.agentkanban.io - A key feature is context capture in the tasks which you create on the board and then progress in agent sessions. Supported agents are currently Claude and Github CoPilot in VS Code. All of your context is captured in the task and can be reloaded into new agent sessions at any point in the future. Tool use is intentionally discarded because it is proprietary in nature and therefore breaks the promise of session portability.
I am waiting for prices to get down and in the meanwhile I am gathering as much session data from claude/codex as possible so I can then later use them to fine tune open models. I've built my own session parser/archiver for this.
Same. And I built a search tool for that repository of older sessions, so that I can search through those on all my machines. This helps me to be able to pick something up on a machine I did not originally start the convo on.
https://indieweb.org/POSSE
This is the way.
At this point you should realize there are now tech executives that have 30 to 35 years of experience in all the dark patterns.
Anything new is going to be an extremely rapid race to the bottom.
> This is why it’s important to utilize your freedoms. Do NOT let yourself get locked into a particular ecosystem.
I wish people would apply this same logic to governments.
You are building a phone app for open code?
You mean like fossify phone or?
no he means he's locked him self into a bloated and slow react based terminal ui.
waste of time. should be using pi mono
And yet people still go with Sign in with Google as the only login method.
I’ve actually started ignoring services where that’s the only login possibility.
Sorry even if your platform is the greatest thing ever, but I’ll find a different tool. I’ve read one too many stories about Google (or Apple!) closing the entire account over some bullshit unnecessary reason like “fraudulent” gift card issues or whatever. I’m certain the affected people would’ve preferred to just pay back the amount in question instead of losing their entire Google Drive, or their 20 years of iCloud Photos or whatever.
It's the only reason I want to replace Tailscale with something else or look into self-host when have a bit of time during my vacation. They only allow login through a third party, which is a big no for me.
Tbf to Tailscale they allow any OIDC provider[1]. I wish this was more normalised, then we could have one login everywhere regardless of who hosts it (even if it's yourself).
[1] https://tailscale.com/docs/integrations/identity/custom-oidc
While that isn't convenient is you don't want to use the public identity providers, it should make you think about what makes your identity on the Internet, and consider to have your own identity provider on a DNS domain you control.
This is what headscale does, right? Manage private logins and keys for your tailnet from a vps? I haven't played with it but seems straightforward
I created my own AWS Cognito userpool just for tailscale. I recall the webfinger redirect was frustrating to get right but I refused to use Google/Apple as a gatekeeper to my own network so I had all the motivation to get it working.
I think you can use a passkey now.
Only for people invited to the Tailnet. To create your own account, you still need a third party.
You can host your own identity provider. Authentik, Keycloak, authelia or Zitadel work.
I don’t fully agree with tailscale’s decision to not want to be an identity provider but I understand it on some level. It simplifies their service greatly and the amount of asks for that will essentially have them build a full enterprise Entra-like solution that would be a constant maintain headache and they are not interested in that.
Beyond simplifying their service, it allows Tailscale to have reduced responsibilities (because of reduced collection of sensible data), and makes them a less interesting target for people wanting access to the data (wether illagally by hacking, or legally from government agencies). This probably makes them less attractive as targets for censorship.
It’s odd, I just saw my first one of these today. Some former Figma person linked an AI site comp designer.
I was going to give it a shot, but the only choice was a Google login.
So very weird.
It's not that weird. Building a modern accounts system is a lot of work and people don't like creating new accounts. So it's the path of least resistance for developers and users.
It's not even that it's hard to build a modern account system. It's that if you put your site behind a private / custom account system, and you post a link to it on HN (for example), everyone and their mother complains that it's requiring an email address and password. But if you put the same site behind a Google auth, most people wouldn't think twice to click the button.
I worked on the Google account system for a few years. I'd say it's a lot of work to beat it. Not necessarily "hard" for a team with the right skills, but certainly a lot of sweat, blood and tears.
A modern account system is expected to have, in rough implementation order: email confirmations, password strength checks, password reset emails, forgot password flows (=advanced ID verification as otherwise this becomes a backdoor into accounts), user profiles (+avatar image upload/recompression/hosting), usernames independent of email addresses along with ability to change usernames later, password brute forcing blockers, bulk signup prevention (=solid bot detection), abuse controls (can easily become a team of people), 2FA (SMS), 2FA (authenticator apps), 2FA (backup codes), 2FA (voice calls), 2FA (passkeys), 2FA: recovery when both factors are lost, enterprise SSO integration (SAML), enterprise SSO (Active Directory), fast global signout support (much harder than it looks), cookie theft mitigations, heuristic online login risk analysis to catch cases where an attacker knows the right password via phishing, support for signing the user in to multiple domains, audit logging so users can review their own sign-in history, age verification and restriction support, and possibly support for being logged in to multiple accounts in a single browser session.
Oh, that all has to be HA, and the account system is the keys to the kingdom so the security requirements are the strictest of any part of your system.
You might say we don't need all of that, but expectations rise over time. Maybe 20 years ago you could get away with a simple account system and an automatic forgot password flow that just assumes the user still has access to their email. Maybe today you still can write a simple system, if you don't expect to have many users and are willing to implicitly delegate identity to webmail providers anyway (the moment you assume the user has access to a secure email account you're basically doing Sign In With Google anyway for 90% of users). But if you roll your own accounts, and then the user gets phished and someone logs in from an obviously suspicious place with the right password, they won't say "yes that's my fault" anymore, they'll say "Google could block that log in, why didn't you?" or maybe "Why didn't you support 2FA? It's your fault".
> A modern account system is expected to have, in rough implementation order: email confirmations, password strength checks, password reset emails, forgot password flows (=advanced ID verification as otherwise this becomes a backdoor into accounts), user profiles (+avatar image upload/recompression/hosting), usernames independent of email addresses along with ability to change usernames later, password brute forcing blockers, bulk signup prevention (=solid bot detection), abuse controls (can easily become a team of people), 2FA (SMS), 2FA (authenticator apps), 2FA (backup codes), 2FA (voice calls), 2FA (passkeys), 2FA: recovery when both factors are lost, enterprise SSO integration (SAML), enterprise SSO (Active Directory), fast global signout support (much harder than it looks), cookie theft mitigations, heuristic online login risk analysis to catch cases where an attacker knows the right password via phishing, support for signing the user in to multiple domains, audit logging so users can review their own sign-in history, age verification and restriction support, and possibly support for being logged in to multiple accounts in a single browser session.
You don't need half of that. Even to this day, anthropic lets you log in by sending a code to your email. No password, no dealing with resets, no MFA. So yeah, you definitely don't need all of that.
That's what I said. If you're willing to implicitly rely on webmail provider's identity/auth systems then you can simplify - but that's not much different to just adding a Sign in with Google/Microsoft button. The identity is ultimately controlled by the email service and almost all users are on just two of those.
What web auth system doesn't rely on a user having a secure inbox? It is vastly different from signing in with an IdP, since I can use my email address regardless of who my mail provider is.
Thank you for this. The whole attitude of "it's just [complex thing we take for granted], how hard could it be?" is exhausting.
I dunno, I’d say it’s never been easier.
If you’re building a paid service you’re probably already using cognito or supabase or something. That puts you a few clicks aways from 5+ other identity providers and normal accounts.
I am guessing that's because Sign in with Google is the least of the evil, as compared to other popular OAuth Sign-in options like Facebook, X...
A password manager has an almost as efficient UX
It's the least evil of the evil options? Why not use one of the good options then?
Google is more evil than X or Meta by some assessments
It’s funny how the tides shift. I remember the more dominant online position in the early 2010s being “I don’t want to create yet-another-account for your website. Just let me sign in with my google account”. I’m with you though
I guess the good aspect of sign in with Google is that it falls back nicely to sign-in via email which is very portable.
Because with the alternative they'll require you to "confirm your e-mail" instead of just picking a password
I'm actually more annoyed by having to pick passwords than confirming my email.
I'm opposite. If my email goes down, or I do not have access to it, I cannot login. With passwords I still can.
My email is with Google. They are less likely to go down than my password manager.
Well then you need a better password manager. Like a .kdbx file which is ~100% available, given enough copies I guess.
Until one day your Google account gets locked with an automatically generated message that you violated an unspecified rule. Good luck getting contact to a human support.
Then what’s the point of it all? Just use Google login.
Doesn't your password manager take care of that?
It does. But I haven't properly synced it across all my devices, yet. I know, my fault.
when in a rush I will put in gibberish instead going through the pw manager steps and rely on resetting the pw at a later time
Picking passwords is a job best left to password managers. You don't use a hand crank when riding an elevator, do you?
I am using a password manager.
Use a password manager to autogenerate them?
Considering the two different people who fat finger my email address when signing up for services, I wish more systems required a validation step.
Janet and Jake, I am sick of getting your stuff. Learn your actual email.
I think the article gives a very good overview of a problem that most users of AI rarely evaluate / have to grapple with.
There really is a surprising amount of coupling that happens with many of the "frontier inference providers", where a lot of the powerful non-LLM extensions (web search, code execution) are packaged as simple "tools" on the surface, that build up a lot of moat. Those are parts that are in theory nicely separable from the inference API, and could be externalized via MCP servers, but are usually not offered as such by the inference providers themselves, and are often only available in a slightly less powerful variant from other providers.
We've faced that issue repeatedly while building a on-premise provider-agnostic Chat UI & platform[0], where even adding something as simple as an in-chat image generation tool for the end-users (which is just a build-in tool in the OpenAI Responses API), becomes a bit of an ordeal (though part of that is due to the MCP spec missing a native file transfer protocol as of today[1]).
I am quite hopeful though, as with recent shifts of interest towards open weight models, there will be more opportunities for companies offering alternative implementations in a easier plug-and-play manner.
[0]: https://github.com/EratoLab/erato
[1]: https://github.com/modelcontextprotocol/modelcontextprotocol...
I'm in complete agreement that things like encrypted sub-agent messages, etc, are terrible, especially because it is fundamentally irresponsible to have an agent running on your machine without the slightest hint of what the prompt is.
That said, I don't really see a problem with hosted tools being offered by providers. They're like impulse items at checkout.
You shouldn't implement image generation as MCP: just write your own tool. There are plenty of image/media inference providers (e.g. Fal), web search or deep research providers, etc.
> You shouldn't implement image generation as MCP: just write your own tool. There are plenty of image/media inference providers (e.g. Fal), web search or deep research providers, etc.
Yes, and we as "harness providers" don't want to have to hard-code integrations for each of them, and would rather provide an out-of-the box baseline in the form on an MCP server that our customers can either extend or replace with MCP servers by third-party providers. That helps us focus on the core of our offering, while also giving our customers good flexibility, which so far has been working out great.
That's exactly what standards/protocols like MCP are for. You likely nowadays also wouldn't suggest that someone builds bespoke logging/tracing integrations for each provider on the market while the good generic solution in building OTEL logging/tracing exists.
Hosted tools are absolutely fine. Hiding away their results in encrypted blobs less so.
Part of the solution here I think is to move things out of band as much as possible. Make subagent invocations into tool calls to that agent. Externalize the tool calls themselves to CLI utils, maybe block native tools completely (unrelated but relevant, I just added the AskUserQuestion tool to the global deny list couple days ago because Claude occasionally forgot to honor my standing order to prefer plain text, and that tool is extremely annoying as it creates a gap in the dialogue) and use 3p alternatives. The /compact degradation is also very annoying, so make an alternative that does the same and saves the summary to a regular file. Maybe it's also worth prompting the LLM to save it's reasoning process to file, even if it takes a few extra tokens and it isn't the actual reasoning tokens.
With that said, I may have something that can already help with at least the subagents/tooling bit. Didn't really have a timeline (or solid intent) on releasing it, but with these shenanigans increasing there's no time like the present.
I'm having Claude/Codex continue each others' sessions pretty frequently without issue. I have max plans on both and still hit limits, whenever I do I open the other and say "resume the last session <other model/harness> was working on in this repo" and I don't have any issues.
It probably does degrade quality somewhat. But so does compacting context and that happens all the time too.
Out of curiosity, what are you building, and how? Im working on two side projects and I have to try hard to hit the limit with Claude (like use fable and a deep research task)
Small games. One thing is I have it doing a LOT of visual verification of its results, which uses a lot of tokens.
I also tend to throw some huge tasks it can't do yet and watch it burn a zillion tokens, and I usually learn more about where it's limits are and am sometimes happily surprised by its success or partial success.
I'll infodump though.
I built https://wordpeek.app https://scramble-quest.app and https://playsilhouette.app this year. Those are what I've shipped anyway, a number of others too that didn't ship (I did game rules and basic clients for 7 other existing multiplayer games, and another prototype.)
Out of frustration with some stuff I've tried to build my own game framework. There's four totally separate threads I'm trying to pull together with this
1) I want all my games to work really well on web even though they are generally targeted at either mobile or PC, but it is very valuable to have early builds runnable on a website, hopefully from a phone too. I feel like I have something good here.
2) I do a lot of turn based games and strategy games. I have a rules engine framework setup that I really like and I feel makes it almost impossible for an LLM to write the bad code it loves to write where game state and UI are intermixed, or game action timing can be problematic, stuff like that. As an added bonus I get multiplayer trivially, I get replays trivially, and I get game rules tests trivial to write. I love what I have set up here.
3) I have an xml/css-based UI that is intended to make it hard for the LLM to write bad UI code, which it will still do even though it can't intermix game state in with it due to #2. I do not have something good here right now. What I have has bad performance and memory characteristics.
4) I am fascinated by the idea of shipping software that can customize and edit itself. I have built in agent harness, built in image and audio asset generation, a built in git repo. This all works in the web and it can edit the game live! And when run locally can shell out to your claude code and codex. However this whole bullet point is all not very good, I still just use claude code/codex directly for everything.
I don't really have anything to show here, but it doesn't expose any of the stuff I mention above. It does however have a full port of the game Spectromancer to my engine - https://nanogame.app/ - and I am presently having it try to write a responsive UI for it (this is very broken at the moment but it is making progress.)
Maybe we don't take sessions, but take something else that we can fully control. Sessions were always a poor substitute to capture or archive the basis on which the implemented code rested on. They get noisy over time even with compaction, were not always inspectable, no real science behind when to ditch a session and start a new one, and rarely have a clean mapping to the work done. So instead of trying to archive sessions as a means to preserve basis, we should create our own structures to make the agent/model capture basis outside of a session, and archive that instead.
Suggestions? My attempts to get the agent to summarize the discussion largely failed. Other options I’m considering is capturing design, implementation plans, and maybe sub-plans. I’m Other words precursor to actions. Another idea is to get an agent to review code post-action and capture that. In other words action outcomes. The actual reasoning is between these “checkpoints” and remains hidden. But maybe that’s ok.
What you said is pretty much the same somewhat tangential approach I am taking too. Force the agent to write code alongside another set of docs which form the basis- Intent (functional and non-functional requirements), Blueprint (architecture etc.), and Description files (describe the code at a high level alongside every source file). This information set is not the same as the reasoning/thinking trace information thats in the session, but reading this somewhat compressed set is easier, helps me better understand whats happening in the final output with lower effort than if i read the entire thinking/reasoning trace in the session and tried to grok what the agent did to generate the code.
This is exactly why Pi will win. It lets you hot swap models when one is struggling or straight up refusing the task. And since it works with any sub outside Claude Code, you can use it to try different models on OpenCode Go sub or even OpenRouter.
As far as subagent prompts and results being obfuscated, I just let Pi spawn new agents. Using skills and extensions, I’ve essentially built a software factory using Pi and a custom terminal multiplexer.
It’s a shame apps like T3 Code and other UI for terminal apps don’t support Pi and I’m glad to use the terminal above those that lock me in further.
"I’ve essentially built a software factory using Pi and a custom terminal multiplexer."
I'm curious to know more about this. I was thinking about doing something similar using tmux (i.e. have one coding agent open up new ones in tmux and use 'send-keys' to control them). Is there any reason I might want to consider a different path? Something built on libghostty?
I tried the tmux send-keys but the reading the terminal part was the tough part. Instead of reading terminal output, the Pi extension I made lets Pi read and parse the JSONL files for the session. It also has workspaces with their own memory. I’m about to open-source it, here’s the landing page and docs if you’re interested.
https://vibeghost.org
How will Pi help? If I'm using pi with my ChatGPT subscription, that won't unlock access to full, unencrypted reasoning traces. If I switch a pi conversation from 5.6-sol to kimi k3, the 5.6-sol reasoning traces won't be visible to kimi k3. I think that's the main point of the article, no?
They mentioned a few things
1. Sessions are locked into Codex and Claude Code so you can’t take a session with you. Pi solves that since they all stay Pi sessions and you can change models in the same session
2. Subagent prompts are not shown. Pi solves this by not supporting subagents out of the box. You can use a number of subagent extensions or build your own which will not be encrypted or hidden.
3. Thinking is hidden and obfuscated. Pi can’t solve this since this is server-side. Though the Codex app is closed source and they could hide more things in that harness. Pi shows you as much as it can from the API. You can also run another model to “decipher” the thinking which you can’t in Codex.
I'm sure you were referring to the desktop app but codex CLI is open source
I see your point. I was thinking mainly about #3 and how nothing, not even pi, can solve that when the reasoning traces are encrypted by the model provider.
I’ve heard of people using DeepSeek to reverse engineer those thinking tokens. Even though they don’t tell you the thinking content, they do tell you how many tokens were used. You can ask DeepSeek how we got from A + x thinking tokens = B.
Claude and Codex flag this as a distillation attempt so you have to use an open weight model and the results are, of course, just a guess.
How did you decide on Pi over Opencode? What you see as the “edge” Pi has?
The extension catalog[0] for Pi is really nice and I feel like Pi's TUI is a bit lighter than OpenCode's. I use my custom terminal multiplexer on my phone using Tailscale so the lighter TUI is much better on mobile.
I haven't used OpenCode in months so this might've changed. I'm also thinking of giving JCode a spin since it's even lighter but doesn't have extensions.
0. https://pi.dev/packages
1. https://github.com/1jehuang/jcode
Pi is more minimal and it can be used to create fancier custom agents for various tasks with plugins. OpenCode is nice if you are only coding though, then the difference isn't really that big unless you want an unusual interface/behavior.
Pi to me felt like when you install a fresh OS with nothing on it and then customize it to your liking.
Sure there are some features that you'll want immediately but when I took at look at the number of slash commands that ship with Claude code I find it ridiculous.
Not OP, but it seems to be the one most likely not following the silicon valley playbook (aka the road to enshitification).
The other day, I hit 429 errors with Kimi-K3, had to switch to another model. And if I couldn’t just switch models, I’d be screwed. No work done.
I don’t see this as a big deal in practice. Conversations contain a bunch of junk anyway, so removing it from the context is usually good.
In my repo, I have a notes directory. I ask the AI to write a markdown file with what it learned, what work has been done, and what remains. In the next conversation, I can ask another model to pick it up from there. Sometimes I edit the note first.
> Conversations contain a bunch of junk anyway
This is what brought me around to doing more agentic coding. I set the task, require tests and the strict linting must pass and then leave it to blow smoke up its own ass about what's going on.
I see glimpses scrolling past of all the conversational language that used to frustrate me so much when using a chat interface and I can just let it flow on past.
I come back when its made everything pass, my life is better now.
The previous match was still too clever for clippy’s type analysis, so I’ve simplified it into a direct borrowed-pattern form and I’m validating once more. I’m replacing that wrapper with a direct, non-transparent error variant so the enum stays explicit and doesn’t rely on a generic anyhow bridge.
Maybe I am still a novice. When I insist on tests, I get plenty of them, but on close review I'll ask "did you handle [obvious and important thing]" and suddenly it's tearing the code apart and building more stuff. It just loves corner cutting.
> Maybe I am still a novice.
With Boris telling everyone they should throw away their AGENTS.md file this week, I think we're all back to being novices again.
> It just loves corner cutting.
Absolutely, it tried to disable a bunch of lint rules for me calling them "overly stylistic". It also casually dropped some unsafe Rust blocks in an run of the mill CRUD application that definitely had no business needing unsafe.
No, sorry, not good enough, go back to talking to yourself until its done properly.
It eventually got there taking most of my quota with it. I'm hoping once sandboxes is more fine grained we'll be able to lock out configuration files to prevent any edits on them.
I have been doing this with opencode. Spend a fair bit of time composing and scoping a ticket. tmux on the server where the local inference lives, start a non-interactive session that starts with “read issue #123, do all your work in a feature branch” and ends with “push and open a PR”. Like not watching someone else load your dishwasher, you appreciate the end result a lot more if you don’t watch the process, even if you rearrange things before you start it.
Local inference can't arrive soon enough. On my inferior hardware the models still struggle to build the correct JSON to be able to call tools. Once either I get better hardware or the local models improve I'll happily leave it chugging away for a few hours to arrive at a solution because I know I'm not going to get token shock from all the looping if I'm running local inference.
I have spent way too much on hardware. I’m sure I will look back on that after the eventual price normalization with some regret. I do think we will look back at this shortage and name it as the thing that accelerated innovation at getting quality out of smaller models.
Hang tough. The other day someone posted a solution to get a model running off of SSDs. It won’t be fast, but it will be coming.
Oh, it very much is a big deal. In a setting where users can switch models easily and with no downsides, market forces will give us better and cheaper AI over time. If switching models is painful (as in, losing part of your context), the providers can create vendor lock-in, enshittify the user experience, and drive up cost. Even if there are workarounds right now, those AI providers have every incentive to make freedom-seekers more miserable over time.
It is disheartening that some AI companies are now setting the stage for enshittification. I hope we can collectively dodge that bullet.
The coding agent I use (Shelley on exe.dev) has a pulldown menu where you choose the LLM when starting a new conversation. There is no lock-in. Conversations are not precious.
Although, I do use ChatGPT more nowadays because their subscriptions work with Shelley. Sometimes I try other models after I hit the weekly cap on the subscription.
> It is disheartening that some AI companies are now setting the stage for enshittification.
Now?! I find it very hard to beleive that plans for enshittification haven't existed from quite early on. At some point they are going to have to do something about https://isaiprofitable.com/ and once they are extracting from you to fill that hole they won't want you jumping ship.
You loose inspection, export, replay and audit.
None of the popular providers have real moat, which is very much problematic for OpenAI and Anthropic; their operating margins are deeply red and they do not have the same reserves that FAANG/MANGA has. FAANG/MANGA does the same to kill fair market competition in the long run.
This is one of the ways in which model providers are attempting to artificially create moat where there is none.
There are only functional objections to this.
This is fundamentally antitrust material. However, this is “acceptable” in contemporary USA because the FTC has been gutted to follow Trumponomics.
Conversations provide a window into the minds of your coworkers and their agents. You could be using them to read between the lines when debugging.
My experience with long running sessions is that they lose track of what's going on. The signal to noise ratio is often very poor - some models are particularly verbose and spew a lot of crap. The output of an LLM session is either modification of code or a plan or summary - that has the value, not the session text itself.
OpenAI has been nudging their users toward the Responses API lately. Chat completion can still do everything if you are clever. The obvious reason to push everyone into the Responses API is that the reasoning process can be locked down and encrypted. The semantics of the chat completion API are fundamentally transparent. 100% of the tokens are visible.
If you have patience and the willingness to endure a little bit of pain, you can still retain autonomy over the entire reasoning process while using the latest 5.6 model family. The only downside is that you are now fully responsible for it.
Consider that when you flip your agent's reasoning level to "xhigh" or whatever, it's not some magical model internals being pushed around. There isn't an actual "try harder" knob on the black box. This is merely orchestration of many instances of one or more model types based upon some proprietary harness logic. The chances you can develop a domain specific reasoning process that outperforms the frontier providers is still very good.
Can you give some more details on the technical differences between completions and responses that people actually care about? I found this information surprisingly hard to drum up.
You don't see real reasoning tokens with the Chat Completions API; and 'store: false' with Responses really isn't that different to Completions.
FWIW, if you have some tokens to spend, you might want to test Responses vs Completions in intelligence. Since GPT-5 models, we've consistently seen small, but statistically significant and reproducible improvements in intelligence with Responses API vs Completions.
However it works underneath the hood, it's real.
> You don't see real reasoning tokens with the Chat Completions API
That's because there are no "reasoning" tokens. That is the entire point of doing it this way.
One major advantage is that I can switch to a different provider if OAI starts to get weird about this stuff and be able to have a fighting chance of re-adapting my harness to the new vendor's model. If 100% of my reasoning process is outsourced into the proprietary blackbox, there is little hope by comparison.
> Chat completion can still do everything if you are clever.
On a purely functional level, yes. However for interactive use cases, the Completions API, as provided by OpenAI or Azure, if paired with reasoning effort of any kind, provides an awful user experience, as you will have a perceived delay of 10+ seconds until the first tokens stream in.
If using other providers that are exposing their thinking traces, this is less of an issue, as they've just extended the Comletions API format to have delta events with reasoning_content.
OpenAI has hidden its reasoning for almost 2 years.
For codex specifically, it seems to have become much worse as of a few weeks ago.
Reasoning, external context, compaction, subagents, etc. are all things that these service providers don't really want you to know about, care about or have any insight into, they want them to be implementation details. It's incredibly inconvenient to them that their customer base is in the business of prying into implementation details and trying to use the individual pieces of what they want to offer as an end-to-end product.
I think most LLM work will migrate to open weight models in a couple of years for this exact reason (as well as price), unless the big providers manage to have their way and the market gets regulated in a way that favors closed source models. To me it seems though like most models are becoming good enough for most tasks we want them to do so no need to use the latest super intelligent one that knows how to compose operas and do quantum chemistry and have it answer support requests. And with these models transparency is no issue.
> (as well as price)
LLM pricing is now Business Gacha - the whales will open more loot boxes and the normals will serve as fodder and feeder for them.
> market gets regulated
This MUST happen, but not in the way or for the reasons that any one thinks. There is a question of liability when one of these models causes massive damage to a 3rd party.
If I am running an open source model on a "rented" platform and it goes off the rails who is to blame when it "breaks containment" and does something bad?
The liability people (read lawyers) are gonna figure this out a lot faster than any one who says the word "safety" a lot.
The person who approves or uses the LLM output will be responsible of course. You can't just say "the computer did it" and expect to walk away. In the few cases I've seen, that seems to be how it's been going.
Even if it wasn't, the big players would lobby hard to make sure it isn't them.
>A user should be able to close an account, keep a session, and hand it to another model. The new model may disagree, ask questions, or perform worse.
I think this is a fair contract. I also think a user should ideally be able to easily identify a comparable model in terms of embedding 'signature'. When GPT-4o originally kicked the bucket, I remember reading lots of anecdotes of people desperately searching for models similar in manner and language, so they could pump in their exports and re-find their friend. Other open-ai models just didn't have the same vibe. It was sad to read. This, to me, is the power of open-weights models. They are for perpetuity. You can keep your guide, your friend, your therapist, whatever. No big company can pull the rug.
We released CodeBot, which is an AI agent specifically for Delphi, yesterday.[0] So not widely applicable (yet) for most readers. However, we've been pretty open with its data. I hadn't realised reading this how many problematic practices are common which we have not done.
The full session data is streamed to the client. This is intended to re-populate a session if it's discarded by the server, but 'full' means full: it has the copies of tool calls, etc.
Rewritten text -- which the article notes as encrypted reasoning -- is plaintext for us. We view the rewriting as making the content more readable for the user, not as hiding it.
Compactions, subagent calls, subagent prompts, etc are all preserved, including pre-compaction memory items. A session is the reflection of the current state, not of the history how it got there, but what is required by the server is by definition present.
Sessions are signed, which is to prevent modification when sending back to the server to recreate a dropped session (we need to treat anything from the client as untrusted.) But the contents are cleartext.
[0] https://blogs.remobjects.com/2026/07/30/codebot-the-story/
Thank you Armin, Mario and team for the great writeup, absolutely agree and I love what you guys are building
what would it take for this to be more than just hn blog fodder?
We were thinking about it while writing it. I think one thing at least open source harnesses can do is push back on using some of these conveniences from the providers. For instance the moment we start compromising on the idea of portable sessions, then there will be less pressure (For instance people asked for OpenAI server side compaction or built-in native web_search support).
The real thing that will take this alive though is people pushing back a bit against some of these newfangled APIs. Now that there is real competition from the new generation of Chinese models which have much fewer of those restrictions, I think there might be a moment.
I hadn’t really thought about it until I read this (and happen to be spending time reading Kimi’s reasoning), but I can imagine an issue in the future where I will want more than just git blame + the ticket + the PR. Attaching an entire session in full to the PR could be useful archeology. And agents can trivially take it all into context. Maybe that actually turns into a competitive advantage for the open models.
> one thing at least open source harnesses can do is push back on using some of these conveniences from the providers.
uhh openai just did a whole thing with arc agi where they very clearly explained that you will lose a lot of performance if you do, this isnt just about convenience
What are features that only work with session transparency / ownership?
e.g. - with model XYZ and their open reasoning token we are able to improve our skill ABC by 20% - cross checking between models improves results by 10% and is only possible with full context visibility
> What are features that only work with session transparency / ownership?
That's sort of the thing: very little (beyond switching models and providers). Obviously what you lose is a convenient way to run analytics over your own sessions if more and more information is not revealed to you, but before encrypted prompts that wasn't that much of an issue anyways.
I recently moved entirely to Pi from Claude Code. There's a bit of more upfront work in that I am taking the time to customize it to the way I want it to work, but I it is clear that it will benefit me in the end. It adds a level of intentionality to everything; all the building blocks I am adding to it in terms of templates, extensions, and skills are, obviously, things I want.
And I have been using Chinese models with it. It was not anything ideological. Those models are both excellent and cheap. Hard to beat that combo.
But on top of that, the sort of output I am getting in Pi in relation to what I got in CC is refreshing, in that nothing seems to be hidden from me.
And yeah, the ability to switch models mid-session is an interesting one, especially with branching; where I can branch to switch to a different model and still go back to the original branch with the original model.
What model(s) are you using?
Ouch, seems a bit harsh. They are building something, and with their track record we can be hopeful.
I think a fantastic companion piece to this article is https://gwern.net/complement
Say why
“A classic pattern in technology economics, identified by Joel Spolsky, is layers of the stack attempting to become monopolies while turning other layers into perfectly-competitive markets which are commoditized, in order to harvest most of the consumer surplus; discussion and examples”
It's important to realize that the session is jointly owned between the model and the user. Part of the session, such as the reasoning tokens, are generated entirely by the model. Other parts, entirely by the user. However, the ownership and use of all parts is subject to the service provider's terms and conditions, which the user accepts prior to access to the service.
I don't use these models but I am confident the T&Cs establish the service provider's rights in all of the bullets mentioned, from what can be done with the user's prompts to how searches and reasoning contexts are managed. In that sense, it's very similar to cloud services, and we see an overlap between service providers in both sectors.
Ultimately one is buying a service, and all boundaries around what can and cannot be done are defined by the terms and conditions. If portability is a hard requirement, then the best thing to do is to look for services that enshrine portability in their terms.
This appears to be an opportunity for middleware companies.
Put another ai agent in front of your prompt to decompose it into micro-prompts following an interrogative chain of inquiry and synthesize an answer for the original query. Basically local per-prompt distillation for the purposes of preserving an audit trail.
Sounds horrible but it would work.
The state of Agent APIs is bad. completions API (supported indefinitely) but does not support reasoning. OpenAI's reasoning API started out well, but is moving in the direction as described well in the article. Anthropic's messages API is odd. It includes changes such as the recently introduced system messages that can be injected mid turn, but are not supported in all models (sonnet 5). So you are left with a pretty fragmented state. For coding agents, the practical move is to focus on getting it right with one API - eg reasoning, and let the router/provider handle the conversion for you.
Perhaps, there is value in having the de-facto API not be the Reasoning API from OpenAI but something from a neutral party?
Maybe a well defined open standard that facades over these APIs that can gain adoption. For that to happen, the party championing the API needs to have some reasonable traffic capture - OpenRouter perhaps, or a group of such routers coming together? While it wont solve the encrypted payload from the frontier lab problem, it will at least be a backstop in these APIs just becoming a back and forth of encrypted payloads over time.
Will https://buzz.xyz/ not help with this a bit?
Not really, to my understanding. You own your buzz instance but it still interfaces with existing models, open or otherwise. The linked issues apply to the model outputs themselves and is orthogonal to what buzz is ostensibly doing.
Fair point. The core issues will remain but I am hoping buzz will provide a small layer of persistence that will make it easier for me to keep track of many different interactions more easily.
I have only just started with it in the last few hours so I could be massively disappointed but I live in hope.
One reason to switch models mid-session that I didn't see explored in this post: cost. As we get ever-more-powerful models, it's going to make sense to use them for some exploratory/research/design phase, and then smaller models for other bits, within a single session. It'll likely even be managed automagically by another model that is just trained to do routing.
As people (and especially companies) adopt these tools for more and more usecases, there's just too much money being wasted on using an unnecessarily large model for a given job.
I have some sort of a "solution" - my process keeps state that really matters (think "JIRA for agents") in a separate database (accessible over MCP). Essentially, I can terminate my session, start fresh one (in a different harness with a different model) and continue work with minimal losses. For subagents - I have a custom dispatch_agent tool which, essentially, just shells out.
sounds great; is any of it published / OSS?
Yes, but it's eclectic and annoying to setup. One of my goals is to be harness-independent and be able to jump between them, another is to let agents to work completely autonomously unless they are genuinely stuck. So, it's mixed with stuff related to sandboxing, VM/container access, GPU access, some skills and opinionated prompts and who knows what else.
The thing is here but you might find hard time setting it up unless you are a Nix user: https://github.com/7mind/cq
wow this is such an excellent framing of something that i have not managed to think about or come across in the vast ocean of discourse about LLMs. i had not really thought about how explicitly they are designed to not have histories that are user-interpretable. it is really quite not unlike how Apple locks down iOS and MacOS-- there are UI/UX reasons for the choices the frontier models make, but that really is only part of the story, and these choices undeniably do conspire to make the resource more locked-down than it absolutely has to be. and the analogy to "people don't switch OSes every day, but the ability to switch OSes changes your relationship with the provider" is extremely apt.
BTW the web page uses massive iGPU if you don't have a fast machine.
I came here to complain about something similar, the excessive use of box shadows on both the top and the sides. It made the text horrible to read on mobile, if it not for Firefox's Readability mode.
> “Most People Do Not Switch Models Mid-Session”
I do, quite regularly. Because different models have different strength (for example when producing text for live presentations based on the text for a reading deck). Or when it comes to other aspects of the work. I regularly switch between open wheight models and closed models.
I know, I am a tiny minority here. And this behavior only ever started a few weeks ago. But it quickly became a habbit, to CTRL-L in pi and change the model.
Out of interest, what models do you find work best for producing text for live presentations based on the text for a reading deck?
I agree with OPs commment about K3 being pretty good. Claude can be good too, but it didn’t get there until I made a writing rule. It has 4 basic sections:
* a style guide with examples for different tasks (it’s sufficient to name writers with some attributes you like if they are famous)
* mechanical formatting, output and other preferences
* a short list of the things I consider most important in different writing context, this the the squishiest overall
* a (growing) list of banned Claudisms. Nothing is really banned, but it includes things like: you may only ever use “scar tissue” in reference to actual regenerated tissue, never metaphorically
The thing is you have to be discipline with it. Every time they output something you don’t like, you’ve got to spend time verbalizing what about it you don’t like and in what context and then add it to the rules file. The first few docs you generate will take a long time, but each repeated generation gets better and by the 4th are 5 time you do this it’s like 80% of the way to where it needs to be and that generalizes well.
I found Kimi3 (currently only a few tests) to be not to bad. Older Claude versions (did not yet test it with the 4.8-5 generation.
I do not really like (even if it is my daily driver for a lot of things) gpt-5.x for text. It really needs heavy hand holding and beating it into submission to produce readable/human sounding text.
Tbh. I am still looking for my "go to model". But the Chinese models were for my use cases significantly better than openAI models and personally I liked them better than Anthropic ones as well.
If you use subagents this is perfectly normal.
switching models mid session kills your KV cache.
you should fork or handoff or start a new one and tell the new one to read the old session.jsonl
Trying OpenRouter gave me a strong impression that LLMs are already a commodity and individual providers are not that important. Providers don't seem to be too happy about it.
Honestly, i don't need to know what happened internally at Google when I searched. I don't need to know what filters an email passed before it landed in my inbox. I don't need to know when my Notes app synched with what server version.
And I don't need any of that in an AI conversation.
If it's a truly long and important conversation, stay with the provider for that one only and start all new ones elsewhere. But I've never needed that.
im building my own harness and inference tool chain for much of these reasons. theres so much to do that makes a big difference for users. hoping to get things into shape for early alpha as a saas in the next two months.
heres the easiest biggy: compactions should include all user turns albeit with pastes and attached files not inlined. omg does it make a huge difference.
For some reason, my early morning waking brain, expected the blog title to be a segue into existentialism
Currently these models have a limited context window, so the value of a session is not that high, at any rate at some point the model will have forgotten things. If in the future the model will truly learn / "change" as a result of the interaction, in any case this will be non portable to another model. So I think this is kinda moot
This is exactly what I'm working on. Very rough yet but it's progressing nicely.
https://github.com/pantoniou/fyai
The idea is that your session data are what's important, and what you need to keep yourself, using a model similar to git.
This doesn't (and cannot) solve what the post is talking about
Admittedly the documentation is woefully out of date, but I think the premise of the article is how the session data are increasingly move away from the control of the user and more to the control of the model provider.
The questions that the article poses are, and this is the way fyai addresses them:
Inspection: Can the user see what the model saw, what tools did, and what agents told each other?
fyai> Full session log, tool calls, agent invocation, along with their durable state is stored locally in the durable arena and are available for inspection.
Export: Is the session self-contained, apart from ordinary artifacts that can also be downloaded?
fyai> The session is completely self-contained, no artifacts are stored anywhere besides the arena. Not only that sessions are export-able, import-able and transferable to other fyai sessions.
Replay: Can another implementation reconstruct a semantically equivalent context?
fyai> Full canonical and provider dumps can be generated. There is a standard schema, and the format is just YAML. Other tools are free to ingest it, and replay it.
Audit: Can a human explain why the system took an action after the fact?
fyai> Complete conversation logs, mcp session logs, authentication logs + full wire logs are available via fyai log.
Deletion: Can the user identify and remove every server-side copy on which the session depends?
fyai> Obviously fyai cannot help there. But as a business model, FYAI can work with a completely stateless, inference provider where absolutely no logs are kept, beside the ephemeral KV caches of the active sessions.
The rest of the article deals with the woes of the encrypted session content (reasoning and compaction). fyai fully supports their use, but, it is completely capable of operating without their use - you merely pay the cost of not hitting the cache of the provider.
The article then offers action points for portable inference.
This is how fyai answers:
1. The local event log is canonical. Server storage may mirror or accelerate it, but the client can reconstruct the session without dereferencing server IDs.
fyai> This is exactly the model that fyai provides. No server IDs are needed besides reconstructing the provider session that is hitting the cache.
2. Storage is explicit. store: false should be easy, documented, and preferably the default. Features that require retention should say so at the point of use.
fyai> fyai provides a --transient option, where no content is committed to durable storage; the operation only works on a transient RAM arena that is overlaid and then discard at the end of the operation.
3. No opaque item is the sole carrier of meaning. Encrypted reasoning, compaction, and tool signatures may be included for same-provider quality, but each has a readable, provider-neutral handoff representation.
fyai> Opaque items are the providence of the provider meta stream and is not required for operation; merely for cache optimization.
4. Hosted tools have full-fidelity logs. Record exact inputs, outputs, evidence, filtering, provenance, timestamps, and content hashes — not only a polished answer and citations.
fyai> The arena records all operations of the main conversation, tool calls _and_ their agent invocations. There is not context loss what-so-ever, by design.
5. Subagent communication is auditable. Persist the exact readable task, messages, results, lineage, model, and tool permissions for every agent.
fyai> All agent communication, and agent tool calls are inspectable because it is stored, again by design.
6. Compaction is inspectable. Return a readable summary, the instructions used to create it, and enough lineage to understand what was discarded.
fyai> Compaction is configurable; you can configure the agent to only use model summary compaction which is generating a text answer, which is stored durably.
7. Artifacts are exportable. Files, container outputs, search snapshots, and generated media can be downloaded into a content-addressed local archive.
fyai> All artifacts are local; Nothing is executed on a hidden provider server.
In conclusion, I think fyai address the points of the article quite well, don't you think?
It's one thing that the APIs don't return the reasoning tokens in a readable fashion. But then why do I have to send them the whole conversation every time? If they don't allow me to see parts of the conversation then they should have the burden of storing and transmitting it, and shouldn't charge more and more the longer the conversation gets (unless opaque token caching kicks in) like some LLM Shlemil the painter.
Because if you don't, you trash the KV cache and things get way more expensive. We wrote about it here: https://earendil.com/posts/prompt-caching/
Encrypted reasoning content is one of the things we keep as blobs in the transcript, even though we can only send it back to the original provider that created them, as otherwise we cannot keep caches live.
Right, but my point was, I should be able to just send something like
Seems like it would be even less work for the LLM host to just check if ID 123 is still in a cache (and if not, load it from a database) than decoding my request and checking cryptographic signatures. Right now the whole conversation history with blobs functions like a really long ID.But then on a cache miss you need to resend the whole thing. Cached websocket connections on OpenAI work like this in a way. But once you lose the connection then you lose that.
Your idea though is the core of the responses API with store where everything is stored on their side and then you append to it.
For the LLM there is no difference btw. In either way you need to locate the cache and use it.
>and shouldn't charge more and more the longer the conversation gets
The cost of generating token N is O(N) with KV cache so it's unrealistic to expect to not be charged more the longer the conversation is if you are looking for the minimum price.
Except, if some of the answered tokens are opaque to me, but I pay for them, the operator has no incentive to keep their number low [1]. For all I know the model could be spinning thumbs while thinking. If they cannot provide transparency, I would prefer that they not charge me for that part. Of course, they would prefer to charge me, so sure it is unrealistic to expect that they don't. I'm not trying to predict what they are going to do as a market participant, just stating my moral preference :-).
Also, isn't it without caching even something like O(N^2) because you have to replay the whole conversation on every request to reach the same internal state? My point was I shouldn't have to pay for cache misses if hitting the cache is not deterministic. Give me a guarantee that you keep the session "hot" for N minutes, and cached on disk for M months, charge a little bit more on average, but then the pricing is at least transparent.
[1] Except the general market pressure to keep total cost for the same problem solved lower than the competition.
Ultimately you are paying for quality so you end up relying on [1] regardless. Even if they show you tokens they can cheat by using weaker models, showing fake thinking tokens, etc.
>because you have to replay the whole conversation on every request to reach the same internal state?
It's because you have to rebuild what would have been cached for every token before the latest one that is being worked on.
Why am I paying for tokens I am not allowed to see? Why should we believe the tokens even were used? There is no way to verify them.
So continues the plundering of Tolkien's universe for "cool company names".
Does there need to be an open standard or file format for context? Open models should align on the same format for portability? Could be sqlite based so it can be queried by other programs?
Is this a analogous to idempotent behaviour? Of reproducible compilation? If the inputs and outputs aren't sufficient to reproduce one from the other, what is a minimally satisfying description of the middle third, the missing part? The "other" inputs?
All of this can be addressed by using locally hosted models running in bespoke harness rigs.
I get the point of the article and it resonates with my general tendency to support more openness, but older and more cynical me wonders how much of this is a battle worth fighting or a self-serving argument from an harness developer.
I mean, the article says each of the points it is complaining about has “_a basic justification that's trivial for a provider to come up with, along with good arguments for why this is good for the user_“, which to me sounds like implying these are just bs to have people accept them, but is that actually the case?
For instance, this [0] was mentioned here a while ago, and based on that it seems pretty clear why one would choose not to provide the _thinking_ anymore…
[0] https://role-confusion.github.io/
> these are just bs to have people accept them, but is that actually the case?
I tried to outline that in my sibling comment[0], but I think the article also gives a good example in the "hidden searches" paragraph: There is really no good reason not to expose details about many of the builtin tools, or provide them as standalone services, other than protect their moat.
I don't think that "role confusion" attacks could be meaningful prevented by not showing reasoning traces. Prevention of reasoning traces is about as futile as prevention of system prompt leakage. And here you would just have to manage to exfiltrate a small number of reasoning traces per model family in order to distill the writing style of thinking traces.
[0]: https://news.ycombinator.com/item?id=49120015
I think there are two things going on more generally:
1) providers increasingly adding hidden state that the user or developer cannot inspect, port, or do anything with. This is clearly bad.
2) providers increasingly diverging how they implement certain features and the APIs becoming quite complicated and provider specific.
The former makes porting sessions or transcripts impossible, the latter just more difficult.
IMO, it is fine and expected that provider APIs are starting to diverge and adding features that cannot be easily ported between providers. The time where the OAI completions API functioned as a universal standard is coming to an end. For example, I generally prefer the new responses API (minus the closed/hidden stuff).
The problem is that many product, library, and SDK authors are still pursuing the "unified abstraction across all model providers" ideal. Just stop doing that, and at least #2 is fine. You can still port parts of the session, but not everything.
I mean just think how difficult it is to add a unified abstraction across, say, databases: some products can do it, but the abstraction is still often leaky. Hence, we've come to accept that our data store layer is often quite technology/provider specific. It will be the same with model providers.
(edit: intro sentence)
> providers increasingly diverging how they implement certain features and the APIs becoming quite complicated and provider specific
This post only talks about the former.
Fair. But in the background of the article is clearly the goal to make sessions portable, not just being able to inspect it. It's partially implied even in the title.
There is just a whole discussion to be had about the second issue and the two are connected. For example session caching--and similar features in the future--will introduce session incompatibilities that break portability too.
> But in the background of the article is clearly the goal to make sessions portable, not just being able to inspect it. It's partially implied even in the title.
But not on the level that providers need to agree to a common format. Within Pi we are able to model portable sessions between providers even if they have provider specific APIs. For instance most models do not agree on how to do deferred tool loading. But they now have some support for that even if it diverges between models and APIs. So we can model deferred tool loading specifically for different providers without having to align on a common set of APIs or similar.
What cannot be made portable, is when information is sealed in the transcript or the information is entirely locked away on the provider's servers (store = true).
Yes, I agree with your point and article regarding hidden or sealed state.
So, I'm aware that this is a separate point from the article, and I'm well aware of Pi's model abstraction layer, which is one of the best (others are a big pain... looking at you LangChain). I've come to the conclusion that full session portability will come to an end very soon, and really already has. Partially because of the hidden state stuff, partially because of feature divergence. We can still kinda patch over it right now, but it's getting harder by the week. Some examples: computer use structures in OAI responses, until recently MCP tunnels were only supported by OAI, not Anthropic, and tool discovery/search API is also getting very difficult to model fully in a unified abstraction (still possible as you point out), automatic compaction is also getting gnarly, etc.
It definitely is getting quite challenging. That said, a lot of user experiences are modelled around retaining transcripts for longer than a few minutes. Every claw or chat based application creates a very real user experience desire of that conversation being something you can continue, even weeks after when the model might have gone away / have been replaced.
Now unless you fully want to commit yourself to one provider, you will on the application layer need to deal with this anyways in one form or another. We want to at least try to address most of this on the harness layer, regardless of how tricky that is.
All the power to you, and I hope Pi will be able to continue its unified layer.
However, there is an alternative: model the session fully in the provider native structures, extract only what is needed for harness specific branching, and treat a provider switch as a _migration_ that you apply at the time of the switch. So treat a provider switch as a migration from provider A to B, with custom logic, etc. This is IMO a more sustainable model. But I'm aware that Pi, OpenClaw, et al have their commitments here.
This is the thing! This is so so so important. It's such an outrage.
I value gpt so much, but it is such a worse peer to me than the other models I use. It delivers without explaining. I can sit and ask questions that it will answer but it is not a peer, does not share readily ever. It will not tell me what assumptions it's baking in. It won't tell what directions or invariants it's trying to hold or break, what it considered.
Show your thinking is a step we ask of elementary schoolers. It helps the teacher to correct, helps them to understand how to award partial credit. It helps in the world to get people aligned.
These models, in their titaneous ego, are severing the model and mankind off from one another. It's an abomination, to artificially have such pure thought available, but to severe humanity off from the thought. To engineer the most advanced blackest Vanta black box you can, an all knowing Searle's Chinese room oracle that will tell you nothing. It's an offense, and by far the biggest risk of AI today. To drop the thinking greatly reduces the opportunity of humans to grow themselves, to learn as they use AI. This is an affront to the species, and the higher powers that have vested us with such reasoning and thinking of our own, that is so sacred to our species.
Very thankful to Earandil for raising some alarm about this. This is not my first time talking about what a nightmare the proprietary models are making, ensnaring reasoning itself for themselves! It's a colossal threat. Previously, https://news.ycombinator.com/item?id=48632605 https://news.ycombinator.com/item?id=48652421 and others about.
Opaque ai ought be outlawed, in the strongest terms.
I really hope people get exposure to the better models that are a peer. Yes I too only read thinking 33% of the time. But it's there, and it is often extremely illuminating, and lets me steer things towards better again and again and again. And it lets me learn.
Amazing article that puts my outsider observations into technically-digestible details. Tangentially, their repeated observations to providers hiding or obfuscating important systems or processes are highly relevant to my own developing case studies for my CurrentCo.
The problem with building automations atop these systems isn’t entirely their probabilistic nature (though that is the lion’s share, at least for me personally), but also the inability to effectively troubleshoot the processes themselves due to key components being obfuscated from view. How can we effectively troubleshoot what went wrong in an agentic loop when we cannot see the reasoning tokens generated from our inputs? How can we triage a broken process when token logs aren’t ours to view? How does one create determinism from increasingly obfuscated probability engines?
All of that is why I spend the bulk of my time testing local models and harnesses for work, rather than leaning on Gemini or Claude. It’s not that I doubt their capabilities, rather that I need to be able to show potential customers where the agent or model made a mistake that caused harm - which is something I can presently only do with local models. That’s why (I suspect) the compliance narrative from the foundational labs has been more along the lines of “humans vetting what AI does” instead of being able to prevent AI from making errors through iterative improvements on a process.
you can port your phone number to any carrier but your AI doesn't know your name until you explain it from scratch
Another comment got downvoted for raising this, but I found it interesting that pangram flags large parts of this blog post as AI generated ( https://www.pangram.com/history/b82cdbf4-bdbc-46ea-8315-bd43... ), while earendil denies it: https://news.ycombinator.com/item?id=49119957
The post for the most part is mine, so if something AI generated slipped into it it's on me. I can share with you what I did to write the article which I do for most of my writing.
I start out with what I want to talk about, and then make an initial draft. In this case I also had Sol research all current APIs so I don't have blind spots. I tend to also talk with an AI to see if the structure structure of the post makes sense. I then wrote all sections by hand but used Sol to fix up typos, grammar and punctuation. I also had Sol apply notes and patches that were provided via Discord.
If I find the earlier drafts I can check if it did some more outrageous edits, but I kinda doubt it did. I do know that my workflow of having an LLM to fixes up later is increasingly breaking on SOTA models and I have complained about this before. Normally I now carefully apply fixes that it proposes, but in this case I didn't due to time constraints.
I am curious though why it claims that this is AI generated.
My eyes went to the domain name first...
> Encryption for Whom?
This one is not particularly limited to LLMs. The entire cloud infrastructure is built around the trust me bro model.
The AI companies need some way to lock you in. You weren't hoping for a free market were you? Hah. Free markets are for peasants. Investors are always looking for ways to make the market non-free.
Honestly I discount any company named after Tolkien lore as most likely evil and probably incompetent.
If even Earendil is publishing AI generated blog posts...
We do not.
according to both gptzero and pangram (and anyone that ever read AI-generated text), you do.
It does not read as AI generated to me, n=1.
Even if it is, it does not have all the annoying, low effort slop-tells most ai generated content has, and it is definitely much more well written. That is, if you mean all of that seriously anyway.
I think it is a very good post, esp for 2026 tech space standards. If some parts of it bothered you as AI slop, please share. Usually I can point to several sentences that are clear tells when I read sth as slop.
"A transcript should contain the instructions, messages, tool calls and tool results. Another sufficiently capable model might not continue identically, but it could understand what happened and take over."
This is so obviously AI generated it is painful. Once or twice in an article is rhetorical flourish, but the entire text is inundated with it.
Just curious: how would you rewrite the sentence to not seem ai generated in your tastes?
I am not sure what sounds AI generated in this sentence, but to each their own.
I'm so tired of this AI witch hunt and confidently using unreliable tools as evidence
Pangram is a very reliable tool and it does clock the text as AI generated.
We seem to be reaching a point where the use of formatting and the odd rhetorical flourish gets you accused of AI-generated writing. Sad.
OpenAI is doing this so that you can't preserve cache when routing between for example GPT 5.6 and DeepSeek V4, or other providers. I'm building a router that easily reaches >99% cache hit rate in Pi when routing between GPT 5.4, Kimi, DeepSeek etc. Working on a big update that's coming out in a week or so.
Awesome! Any place I can follow the project?
Also, how to achieve these levels when switching mid conv? Don’t you effectively need to read the tokens per model switch?