Notes · Autonomous agents
The first time my autonomous dev agent ran on its own, it refused to write any code
Best decision it has made. The outage was a broken ruler, and the fix was to the metric, not the feeds.
The first time my autonomous dev agent ran on its own, it refused to write any code. Best decision it's made.
Every Monday an agent works on a side project of mine without me: an AI news board that ranks stories by how many independent sources confirm them. Its loop is boring on purpose:
- Measure the live board.
- Pick one item from a written backlog.
- Build, run the tests, deploy.
- Measure again.
- Roll back if confirmed stories fall 20%, total stories fall 30%, or live sources drop at all.
On its first run, the test suite wasn't reachable from the session. No tests, no proof it wouldn't break something. So it logged “blocked” and went digging instead.
What it found beat any fix. The health check said 5 of 46 news feeds were alive. Most of them were fine.
The metric only counted a timestamp that gets set the first time a story is saved, so any feed that kept seeing the same links looked dead forever.
The outage was a broken ruler. The fix was to the metric, not the feeds. A week later the same check read 43.
Since then: 9 backlog items closed, 0 rollbacks triggered, and 4 tuned numbers the agent is never allowed to touch.
The pitch for autonomous agents is “it ships while you sleep.” I want the one that knows when not to.
What would you need to see before letting an agent deploy with nobody watching?