Flight Recorder replays AI agent failures

This feels genuinely useful because the pain point is not just debugging, it is paying to repeat expensive agent runs after every small fix.

• –The replay model is the right abstraction for agent workflows: preserve what worked, rerun only the broken branch, and keep moving.
• –Root-cause hints matter here because agent failures are usually downstream symptoms, not the actual bug.
• –SQLite plus a CLI lowers adoption friction a lot compared with heavier observability stacks.
• –The scope looks strongest for linear or tree-shaped workflows; tangled shared-state systems will still be harder to reason about.
• –Open source MIT licensing makes it easier to drop into LangChain, CrewAI, or custom agent code without platform lock-in.