Skip to main content
Replay reruns a selected production turn against your agent and shows the first meaningful divergence between the original trace and the replay trace.

​
How to use it

  1. Open Behavior Search.
  2. Click the session or turn you want to replay.
  3. Click Replay turn.
  4. In Reproducibility, choose the Turn if needed.
  5. Enter a public HTTPS endpoint in Target URL (replay_url), or use a saved replay URL for the agent.
  6. Select Replay with earlier conversation context if the replay should include prior turns from the same session.
  7. Set N runs (1-20).
  8. Click Run.
If your agent does not have a replay endpoint yet, use Configure Replay URL or the replay prompt in Set Up with a Coding Agent.

​
How to interpret the results

Kindred compares the original trace to each replay trace and reports whether the run matched, had minor variance, or diverged. The most important field is the first divergence. It identifies the earliest meaningful difference Kindred found, such as a changed tool selection, changed tool arguments, changed tool result, missing step, extra step, different step type, changed assistant output, or changed unstructured content. Common replay errors include:
  • This turn is not replayable (no replay context available). Run is disabled. The selected turn does not include enough replay context.
  • Set a target URL to run reproducibility. Add a Target URL (replay_url) before running Replay.
  • Outbound URL must use public HTTPS. Use a deployed HTTPS endpoint or a public HTTPS tunnel.
  • Outbound URL hostname is blocked or Outbound URL resolves to a blocked private or local address. Localhost and private network URLs are blocked in production.
  • Run failed: The replay endpoint did not return successfully; use the error message shown in the UI.

​
Next steps