The flight recorder for AI agents

Know exactly why your AI agent failed.

AI agents decide and act on their own. When one goes wrong in production, logs won't tell you what it was thinking. Runback replays the agent step by step — so you see the moment it broke, and fix it in minutes instead of days.

research-email-agent · run failed
Research Next.js 16, email me a summary
agent thinks740 tok
web_search
agent thinks876 tok
fetch_url
agent thinks998 tok
send_email
Found it — one click, no log-digging
The model passed an invalid address: uchit[at]gmail.com
The problem

Teams are shipping AI agents they can't debug.

An agent chains together dozens of decisions and tool calls. When the result is wrong, the trail is a wall of logs — and nobody can say which step caused it, or why.

78%
of enterprises are running AI agent pilots
14%
ever reach reliable production use
40%+
of agent projects are predicted to be scrapped by 2027 — for engineering reasons, not model quality
What you get

Not another log viewer.

Runback turns a run into something you can walk through — and act on.

❮❯

See what the model saw

Every step shows the exact instructions, conversation, and tools the model received — the context that drove its decision.

Replay any step — even edited

Re-run a single decision against the original prompt, or change the prompt and a model and compare the outcome side by side.

Jump straight to the failure

A failed run opens focused on the step that broke, with the error and the bad input already in view. No scrolling required.

How it works

Three steps to a debuggable agent.

Instrument

Wrap your agent with withDebugger() — about three lines. Works with the Vercel AI SDK and any model.

Run

As the agent works, Runback captures every reasoning step and tool call — and the full context behind each one.

Step back

Open the run, walk the timeline, inspect any step, and replay it. The whole story, end to end.

The difference

Logs tell you what happened. Runback shows you why.

A normal log viewer

  • Raw text, printed in the order it happened
  • No record of what the model actually saw
  • Tool calls and the decisions behind them aren't linked
  • Read-only — you can't test a fix without a full re-run

Runback

  • A walkable timeline, with the failure marked
  • The exact context window at every step
  • One click from a tool call to the decision that caused it
  • Replay and edit any step to test a fix instantly
Live

Recent runs

Open one to walk through it yourself.

Stop guessing why your agent failed.

Open a real failing run and see the moment it broke for yourself.

Open a failing run →