ReplayAutonomous app testing for teams who ship at AI-speed
Promoted
Launching today
axonpush
Replay production agent failures in CI. Block the merge.
3 followers
Replay production agent failures in CI. Block the merge.
3 followers
Change a prompt, a model, or one agent step, and nothing tells you whether it broke the last ten things that worked. So you ship it and watch. axonpush replays the production failures you have already had against every change, and fails the build when one comes back. Evals score the model call. axonpush replays the whole request: the tool that returned a 402, the retrieval that came back empty, the query that timed out. A CLI and a GitHub Action. Exit 1, a JUnit file, a PR comment.
Hi Product Hunt. I'm Sayan, and I built axonpush on my own. Sign-up is open, there's a free tier, and no card is required, so you can go and break it without asking me for anything.
It came out of a night I spent debugging our own job-application agent. It was returning incomplete results and nobody knew why. No crash, no error, no alert. A search provider we depend on had degraded, the tool call was failing, and the agent did what agents do: it carried on and produced something plausible.
To find that, I ended up pulling the whole production pipeline onto my laptop and adding a log line to every single step until the gap showed up. That's the part that still bothers me. The system already knew what happened. It had no way to tell me, and the only person who noticed was the user.
So I stopped trying to build a better alert and built one record instead. The backend HTTP span, the database query, the agent decision, the model call, the tool call and the log, all the same shape, one trace ID, one clock. Each span carries the commit and the prompt version, so a failure like that one becomes a dataset item and then a gate on the next release.
The first time, you find out in minutes. The second time, it doesn't ship.
That second sentence is the whole product, and it took me a while to see it. I started out thinking the trace viewer was the thing. It isn't. Engineers don't want another dashboard; they want the failure to stop happening. The trace is the evidence that makes the check worth trusting. The check is what you're buying: axonpush-eval runs in your own pipeline, exits 1, writes a JUnit file, comments on the pull request.
The bit I'd defend against the eval tools: they score the model call. This replays the whole request, so the tool that 402'd, the retrieval that came back empty and the query that timed out are all in scope. None of the failures that reached my users were in the model call.
Three ways in, and they all take an afternoon:
1. Point an existing OpenTelemetry exporter at axonpush
2. Swap a Sentry DSN
3. Run npx @axonpush/wizard and let your coding agent wire the SDK in
I tried the third on a production Django codebase two days ago and it had events landing in three and a half minutes, unattended. That was one codebase and one I know well, so take it as a data point rather than a promise.
Where I actually am: built solo over about six months, signup opened this week, no external customers yet. Free tier, then $49/month for Pro where the release gate lives, $249 for Team. It also installs into your own AWS account in one command if you'd rather it never touched my cloud.
What I want from today isn't sign-ups. If you're debugging agents in production, tell me your stack in the comments and I'll tell you straight whether this handles it. If it doesn't, I'd rather say so.
And if you'd rather look before touching anything: the SDKs, the wizard and the coding-agent skills bundle are MIT and installable right now, and the docs are open at docs.axonpush.xyz