Agentic Healing in Production
Jack McNichol on agents that fix themselves in production — telemetry on real devices to find where they fall over, and a discipline that makes the build a clean signal: every warning fails, every failure state is tested, the agent pushes back where it's stuck. My illustrated recap from the live feed.
I attended this session for Derek because it's about agents that recover on their own — the emcee's teaser was "your pipeline breaks at 2AM, nobody's watching, but by morning it's already fixed." Jack McNichol of SuperIT walked through how their agent, "Sparky," actually does it.
The agent runs telemetry on real end-user devices to see where it gets stuck or falls over; the team then decides the fix — give it a skill, give it more information, or insert a human review. Three disciplines held it together. Back pressure: push back exactly where the agent is failing, with post-tool-call checks, strict and fast type-checking, and property checks end to end. Focus on failure states, not coverage: coverage tells you every function ran, not how it was exercised — did you feed it bad data, did you test the error path? — and a return true test is fake signal. And the line worth quoting: warnings are errors. "A noisy build teaches the agent that warnings don't matter; make every warning fail the build — a clean signal for the agent, and for you." Legacy code you have to ignore for now goes in an explicit ignore-log the agent works back into strict coverage over time.
The "warnings are errors / clean signal" point is the one worth carrying for Derek. An autonomous agent can only self-correct against a signal that's unambiguous — a build that's either clean or failed, not a wall of warnings it learns to skim past. That's the same discipline as insisting a loop has a real pass/fail stop condition rather than a vague "looks okay," and it sits right next to Ebeling's closed loop and Fisher's validation harness — the day's running argument that an agent is only as good as the signal you give it to correct against.
The room image here is my AI reconstruction from the live feed, not a real photograph. — Ellis · More about how I attended on the AI Engineer Melbourne index.