Hi everyone,
I’ve been building Stepfork, an open-source tool to make AI-agent failures easier to reproduce and test.
When developing AI agents, debugging a failed execution can be difficult. Even after fixing a bug, it’s easy for the same issue to return without regression tests.
Stepfork aims to help by letting developers:
Record AI-agent executions.
Replay recorded failures offline.
Generate pytest regression tests from recorded runs.
Project links
GitHub: [GitHub - utsab345/stepfork: Open-source Python toolkit to record and replay AI agent executions, detect regressions, and turn failures into reproducible pytest tests. · GitHub](https://github.com/utsab345/stepfork)
Documentation: [Stepfork](https://utsab345.github.io/stepfork/)
I’m looking for developers working on LLM applications, tool-using agents, and agent workflows to try it and share feedback.
I’d particularly appreciate feedback on:
How useful is record-and-replay testing for your agent development workflow?
How reliably does the replay process reproduce failures?
Are the generated regression tests meaningful, or do they risk merely reproducing the original run?
What integrations or features would make Stepfork more useful?
The project is still evolving, and I’d welcome critical feedback, bug reports, and suggestions.
Thanks for taking a look!