3 ms·
Replay debugger for AI agents (fix failures without rerunning everything)
- whitepaper27 7mo agoHi HN, I recently spent ~6 hours debugging a multi-agent workflow. Every time I fixed a bug, I had to re-run the entire pipeline — repeating LLM calls, hitting APIs again, and waiting minutes just to test one change. So I built Flight Recorder. It records agent execution, shows exactly what failed, and lets you replay from that point — skipping everything that already worked. Demo (30 sec): https://github.com/whitepaper27/Flight-Recorder/blob/main/demo.gif https://github.com/whitepaper27/Flight-Recorder/blob/main/de... Example: @fr.register("crm_lookup") def crm_lookup(email): return db.query(email) # slow API call @fr.register("scorer") def score_lead(contact): assert contact is not None # bug here fr.run(pipeline, "unknown@example.com") # later: $ flight-recorder debug last root cause: crm_lookup returned None # fix the bug, then: $ flight-recorder replay last # crm_lookup is cached (not re-run) # scorer runs with your fix # SUCCESS in seconds The key idea: Fix one step. Don’t rerun everything. It’s open source (MIT) and still early — I’d really appreciate feedback from anyone building agent workflows. GitHub: https://github.com/whitepaper27/Flight-Recorder https://github.com/whitepaper27/Flight-Recorder