Reproduce any past run exactly as it happened, step by step, to debug an issue or understand agent behavior without needing to recreate the original conditions manually.
What replay does
Replay takes a completed run's full trace and steps through it in your dashboard or CLI exactly as it originally executed — the same inputs, the same model responses, the same tool calls and their results. This is deterministic: replaying the same run twice produces the identical sequence, since it's reading recorded data, not re-executing live calls against models or tools.
Replaying via CLI
This streams the run's steps to your terminal in real time, at the same pacing they originally occurred, by default.
Controlling replay speed and stepping
Pause after each step to inspect it before continuing:
Replay faster than real time, useful for long-running agents:
Replaying via dashboard
The dashboard's replay view shows the same information visually — a timeline with play/pause/scrub controls, and a live-updating output panel showing exactly what the agent produced at each step. This is often faster for visual debugging than reading raw trace JSON.
Replay and debugging
When investigating a failed or unexpected run, replay lets you see precisely where behavior diverged from what you expected — for example, confirming whether the model made a poor decision, a tool returned bad data, or a guardrail blocked something it shouldn't have.
Replay does not re-execute side effects
Replaying a run does not resend the actual API calls to Stripe, Slack, or any other connected tool — it shows you what was sent and received the first time, without repeating the real-world action. This makes it safe to replay a run involving a refund, an email, or any other side-effecting action as many times as needed during debugging.
Sharing a replay
Generate a shareable link to a specific run's replay, useful for pulling in a teammate to help debug:
