Back to full roadmap
topicadvanced★ Pro
Deterministic Replay
Replay a failed agent run exactly with seed + mocked tools — gold for root cause analysis.
2 hours1 prereqs
When a production agent fails, the "why?" is hard. Deterministic replay:
- Seed model output:
seedparameter (OpenAI) gives a chance to reproduce (not 100% guaranteed) - Tool mock: record tool calls, mock on replay — re-run the agent without hitting real APIs
- Snapshot state: state snapshot before/after each LLM call
- Diff view: original run vs replay side-by-side
Langfuse/LangSmith support this pattern.