How do you recover when an AI agent goes down the wrong path?

The instinct when a run goes wrong is to correct it: point out the error, ask it to fix the approach, continue. This works for a small deviation and fails for a structural one, and the reason is mechanical rather than psychological. Everything produced so far remains present, including the reasoning that led to the wrong path and the artefacts built on it. Subsequent work is done in the presence of all of it, and the correction is one instruction competing against an accumulated body of contrary material.

The reliable move is therefore to restart rather than to redirect, and the thing that makes restarting cheap is having somewhere to restart from. This is where the working practices pay off: a branch, a snapshot, a copy, a state file written at the end of the last good stage. Teams that have these treat a wrong path as a twenty-minute loss. Teams that do not treat it as a decision between accepting compromised work and redoing several hours of it, and they usually accept the compromised work.

Before restarting, extract what was learned. The failed run almost always discovered something real — a constraint nobody knew about, a file that is not where it should be, a case the brief did not cover. Writing that down before discarding the session is what makes the second attempt better rather than merely different. Discarding without extracting is how teams end up running the same wrong path twice.

The corrected brief should name the misunderstanding explicitly rather than adding general caution. If the agent assumed the data was clean, say it is not and say what is wrong with it. If it optimised for the wrong thing, say which thing. A brief that only becomes more emphatic is addressing effort, which was never the problem, while the specific false assumption remains available to be made again.

There is a judgement call about how early to stop, and the error is nearly always stopping too late. The cost of a wrong path grows with everything built on top of it, so the expected cost of stopping a run that turns out to have been fine is small and the expected cost of letting a wrong one continue is not. In practice the signal is usually available early: the agent's description of what it is doing stops matching what you asked for, and that divergence is visible in the first stage output if anyone looks at it.

Finally, treat a wrong path as information about the setup rather than about the run. One agent misreading one brief is noise. The same misreading twice is a defect in the brief, the material or the tools, and it will keep happening until something outside the agent changes. The most useful thing to do after a recovery is to ask which of those three would have prevented it.