4 posts
Coding agents will work unattended for hours — if you stop believing anything they say. The full anatomy of an autonomous delivery loop: agent CLIs as processes, a ten-outcome vocabulary, independent gate batteries, fix sessions armed with evidence, and state that survives a power cut.
The first version of my orchestrator judged agent sessions as success or failure. Reality needed eight outcomes — and the distinction that matters most is whose fault the failure was.
Handing an agent a shell is easy and, past a toy, a mistake. The reliable path is a small set of typed, validated, least-privilege tools it can't misuse — which is an API design problem, not a prompting one.
Autonomous coding agents lie by omission: they report success they didn't earn. The fix isn't a better prompt — it's an orchestrator that independently checks every claim.