Catch The Waste.Prove It On-Chain.
The Watcher polls real pod metrics and the agents' own logs. The Diagnostician weighs that evidence against any declared intent and returns a verdict with its reasoning. The Negotiator prices the waste and proposes a fix - which only a human can approve.
Heading Component
The Watcher polls real pod metrics and the agents' own logs. The Diagnostician weighs that evidence against any declared intent and returns a verdict with its reasoning. The Negotiator prices the waste and proposes a fix - which only a human can approve.
Built to measure, reason, and remember.
A budget cap only asks how much was spent. An agent retrying a failing task forever, or two agents duplicating the same work, sits comfortably under that cap while producing nothing. Autophagy watches whether the work is real.
A dead allocation and a deliberately reserved standby pod are metrically identical — same request, same zero utilisation, same empty log. No threshold separates them. The Diagnostician reads the declared intent and decides, then shows its reasoning.
Cloud cost tools keep waste private to one organisation. Autophagy attests each confirmed incident on a public testnet, so an agent's efficiency history follows it — checkable by whoever is deciding whether to run it next.
What one orchestrator learns about an agent, the next one can read. The record outlives the dashboard that produced it.
Cloud cost tools keep waste private to one organisation. Autophagy attests each confirmed incident on a public testnet, so an agent's efficiency history follows it — checkable by whoever is deciding whether to run it next.
If metrics-server goes quiet or the model is unreachable, Autophagy reports the gap instead of filling it. A fabricated verdict would look exactly like a real one on screen — so it never produces one.
