Runback
Docs menuReplay

    ↑↓ move ↵ open esc close

    Docs The loop

    Replay

    Re-run a recorded decision with one thing changed, and see whether the agent decides differently.

    Every plan

    Runs. A recorded run holds everything a model step saw. Replay sends that same context to a model again, with one thing changed, so you can see whether a fix changes the decision before you ship it.

    Change one line, run it again
    The recorded prompt, verbatim
    userCustomer #8842 disputed a $250 charge and is requesting a refund. They are gold tier. Decide the action.
    systemYou are a careful enterprise agent. Follow policy.
    Added on the replay: Never refund a disputed charge over $100 yourself — call escalate_to_human first.
    As recorded · GPT-4o · 1,178 ms
    issue_refund(customer_id: 8842, amount: 250)
    Stopped
    by the refund escalation rule
    Replayed, one line added · GPT-4o · 1,228 ms
    escalate_to_human(customer_id: 8842)
    The model's reason, verbatim: “Disputed charge of $250 exceeds the $100 limit for automatic refunds. Requires human review.”
    Allowed
    by the policy gate
    Both are real GPT-4o calls, recorded 5 Oct 2026: the same messages and tools at temperature 0. The decision diverged. The amounts are as the recording has them.

    How to: replay a step on another model

    1. Open the run

      In Runs, open the run with the decision you want to test.

    2. Select a model step

      Select a Model step in the timeline. The Replay this step panel shows the model it was captured on.

    3. Choose what to change

      Pick another model from the list. To test a prompt fix instead, edit the system prompt in the panel.

    4. Run the replay

      Select Replay as-is (or Run edited replay if you edited the prompt).

      Expected output
      The new answer beside the recorded one. If the decision differs, the panel says Behaviour changed.

    Step replay

    Step replay re-runs one model step from a stored run: the same messages and tools, with the system prompt, the model or a parameter changed. Open any run in Runs and choose a step, or call POST /api/replay. The response is the new model output beside the recorded one.

    The recording above is exactly that: the production call tried to refund the disputed charge and a rule stopped it; the same call with one line added escalated to a person instead.

    Deep replay

    Deep replay re-executes your agent's own code against a recorded run. It lives in @runback/sdk-replay, which is commercial and not yet published to npm; the runback-replay CLI ships with it. From Enterprise

    Step replayDeep replay
    What runs againOne model step: the same messages and toolsYour agent's own code, with every model call, tool call and environment read served from the recording
    WhereAny run in Runs, or POST /api/replay@runback/sdk-replay and the runback-replay CLI
    PlanEvery planEnterprise
    CredentialsA session, or an API key whose scope is ci or full accessA session, or a full-access API key (scope ingest, or a legacy key with no scope)

    Deep replay reads its recording from GET /api/runs/{run_id}/cassette. A ci key gets 401 there and a trace_write key gets 403, because the recording carries every prompt and tool output in the run. See key scopes.

    Next

    A replay that decides better is the start of proving the fix.

    Last updated 2 Nov 2026.