Silent agent failures: the $40 retry loop

Your AI agent broke at 3am. Your users noticed. You didn't.

That's the tweet we wrote when launching Puse. It wasn't hypothetical — it was Tuesday.

What happened

A support agent hit a rate limit on Anthropic. The retry wrapper backed off, then retried, then retried again. Each attempt traced nowhere useful. By morning, the project had burned 40× its normal daily spend. Support tickets mentioned "slow responses" — not errors.

Why alerts matter

Error-rate alerts would have fired in the first fifteen minutes. Cost-spike alerts would have caught the retry loop even if errors were swallowed. Either beats learning from churn.

Puse exists so the next 3am failure pages someone. One line to instrument. Slack when regressions hit. Self-hostable so your traces stay yours.

Your AI agents fail quietly. Puse makes it loud.