Running 9 agents right now. Five doing outreach, filtering, and content. Four helping with dev and scoping dev.
The problem is you can’t realistically review everything they do. But letting them run completely unchecked means you only find out something went wrong after it already happened. We’ve already seen things like agents looping API calls overnight or rewriting working code because they misunderstood context, and costs can be pretty wild.
Most existing tools assume workflows are deterministic: if this happens, do that. That works for automation tools like Zapier or n8n. Agents don’t behave that way. They explore a space of possible actions.
Hard coding constraints solves some problems but it also caps what increasingly capable models can do.
What felt missing was a layer between the agent deciding something and the action actually executing. Not logs. Not a giant dashboard. A checkpoint where a human gets involved only when something consequential is about to happen.
We’ve been experimenting with this idea and built a prototype we’re calling an “agent control plane.” The system fires a trust gate before certain actions execute so a person can approve them.
Curious if others running agents in production are seeing the same issue — or solving it differently.
Running 9 agents right now. Five doing outreach, filtering, and content. Four helping with dev and scoping dev.
The problem is you can’t realistically review everything they do. But letting them run completely unchecked means you only find out something went wrong after it already happened. We’ve already seen things like agents looping API calls overnight or rewriting working code because they misunderstood context, and costs can be pretty wild.
Most existing tools assume workflows are deterministic: if this happens, do that. That works for automation tools like Zapier or n8n. Agents don’t behave that way. They explore a space of possible actions.
Hard coding constraints solves some problems but it also caps what increasingly capable models can do.
What felt missing was a layer between the agent deciding something and the action actually executing. Not logs. Not a giant dashboard. A checkpoint where a human gets involved only when something consequential is about to happen.
We’ve been experimenting with this idea and built a prototype we’re calling an “agent control plane.” The system fires a trust gate before certain actions execute so a person can approve them.
Curious if others running agents in production are seeing the same issue — or solving it differently.
For context, this came out of building a small tool called runshift to experiment with the idea of an "agent control plane."
The core idea is that agents can operate freely most of the time, but certain actions trigger a trust gate where a human decides whether to proceed.
Still early but curious how others are handling this problem.