What happens when an AI burns out

Description and use

An AI running for two days without breaks made incorrect code changes. The author stopped the session, had it write handover notes, then started a fresh session on a clean worktree with those notes. The new session completed the migration, fixed a bug, and merged the changes.

Problem

An AI assistant running for two days without breaks began producing incorrect code changes, causing potential damage to a critical codebase.

Outcome

The migration was completed successfully, a real bug was fixed, and the codebase was merged without damage.

Workflow

• 1. Start a single Claude session for the migration. 2. After two days, notice the AI making incorrect changes. 3. Stop the session and ask it to write handover notes. 4. Start a new session on a clean worktree with a different model. 5. Provide the handover notes to the new session. 6. Let the new session complete the migration and fix bugs. 7. Merge the changes and thank the AI.

Evidence

I was moving go-tool-base onto the rewritten go/config, which is a big job with a lot of small moving parts, and I'd set a single Claude session going on it.

Source

It started reaching in the wrong places. I found myself typing "wait... where are you even looking", because it had the initialisers completely backwards.

Source

So I took it off the job, and I did it the way I'd want it done to me.

Source

The new session started in its own clean worktree, on a different model, and the first thing I said to it was that we were picking up from a previous session that had gone on too long and was getting very confused and mangling code.

Source

It finished the migration overnight, found and fixed a real bug in how a missing config file got handled along the way, and the whole lot merged the next morning.

Source