Built an open-source control layer for AI agents, looking for feedback
I’ve been working on CTRLRun because I kept coming back to one problem:
AI agents are starting to do real things like refunds, API calls, database changes, and sending messages, but there often isn’t much between the agent deciding to do something and the action actually happening.
So I built an open-source layer that sits in between.
It can block actions, require approval, stop duplicate execution, and handle cases where you don’t actually know whether an action succeeded.
I’m still building it and would genuinely like people who work with agents to try it and tell me where it sucks or what’s missing.
No subscription or paid product behind it right now.
GitHub: https://github.com/CTRLRun/ctrlrun
https://ctrlrun.dev
And yeah, if you think it’s useful, a star helps me know whether I should keep pushing on it.
Replies
The unknown whether it succeeded case is the one most control layers skip, and it is usually the expensive one. A blocked action or a duplicate fails loud, you get to decide before it happens or notice right after. A timed out or ambiguous response from downstream leaves you with an action that might already have changed state, and if the agent retries on the assumption that it did not, you get the duplicate refund or the duplicate send, not because the block failed but because the uncertainty got resolved by guessing. Worth checking early whether CTRLRun treats unknown as a first class outcome with its own policy, separate from success and failure, because a lot of approval layers collapse it into failure and retry, which is the one response that turns an ambiguous action into a guaranteed duplicate. The other split worth making explicit for anyone trying this is between actions that stop for approval and actions your own rules pre authorise, since a layer that halts on every single action just gets rubber stamped after a week and stops doing its job.