Feature · Replay & debug

Replay a failing run from production.

See the exact prompt the model saw, the exact tool response, the exact retry. Hand engineering a precise repro instead of a screenshot and a guess.

How it works.

Every span carries its inputs and outputs — LLM spans carry the prompt and the completion, tool spans carry args and return values, retry spans carry the back-off window and the trigger error. Open a trace and every step is fully reconstructed.

The prompt view shows you what the model actually saw, including the agent scratchpad assembled by the framework. The tool view shows the exact arguments the agent passed and the exact response it got back. The retry view shows the trigger and the back-off so you can tell whether the agent was patient or panicking.

Share a deep link to any span with engineering — no screenshot, no re-typed JSON, no guess. They open the exact step you saw, with the exact inputs and outputs it captured.

What you can do with it.

Open any failing span and see exactly what it captured — the prompt, completion, tool args, tool response, and retry.
Copy the prompt the model actually saw at any step, including the agent scratchpad.
Take the captured inputs into your own environment to reproduce the failure locally.
Share a deep link to any span with engineering — no screenshot, no re-typed JSON.

What a replayable span looks like.

span · tool.process_refund · 190ms · 409 conflict
$ tool.name process_refund
$ tool.args {order_id: 'A-42', amount_cents: 4200}
$ tool.response 409 conflict {error: 'already_refunded'}
$ next.action retry (back-off 720ms)
$ next.outcome 200 OK {refund_id: 'rf_91...'}
$
$ agent.prompt.at_step_5 (click to expand 1,402 tokens)
$ agent.next_completion 'Your refund is on its way.'

Works with.

LangChain
CrewAI
AutoGen
Pydantic AI
MCP
Custom

Stop chasing screenshots. Start replaying runs.

Free tier. No card required. First trace in under five minutes.

Start freeSee pricing