Heal Your Agents, Don't Just Grade Them

Your evals are green, and users are still hitting failures. The evals aren't wrong. They test what you anticipated, and production at millions of sessions a day keeps inventing failures nobody wrote a test for, many visible only as a pattern across sessions.

Detecting those failures continuously is a hard problem in its own right. Every agent fails differently, the failures that matter are rare, and a detector that flags everything is as useless as one that flags nothing. We've been working on this with customers running agents at scale and by running it on our own agents, learning which failures are worth surfacing and which are just noise.

Along the way, three engineering problems kept coming up:

Economics: you can't afford to run a reasoning model over every session, so what do you screen with, and what earns a closer look?

Noise: a thousand flagged sessions are usually a handful of root causes. How do you collapse them and decide which to fix first?

Trust: when a coding agent proposes the fix as a pull request, what has to be true before you merge it?

How do you verify each fix and expand to cover every failure you find so it stays caught?

You'll leave with the tradeoffs that matter when you move from grading agents to healing them, drawn from teams doing it today