When your agent breaks in production, where is the failure actually?

I'm Sayan, I built axonpush, and it's launching here on Friday. It's a release gate for AI agents: it replays production failures you've already had against every change, and blocks the merge when one comes back.

Before Friday I'd rather learn something than promote something, so, one genuine question.

The last time one of your agents broke for a real user, where was the failure? In the model output, or somewhere else: a tool call that returned an error, a retrieval that came back empty, a timeout, a rate limit?

For me it was never the model call. Our job-application agent started returning incomplete results because a search provider had degraded, the tool call was failing, and the agent carried on and produced something plausible. No crash, no alert. The only person who noticed was the user.

That's the whole reason axonpush exists, and I'm trying to work out how common it actually is.

Happy to answer anything about how it works. And "this wouldn't help me, because X" is the most useful reply I could get.