Agent flight recorderOpen the recorder

Every agent run,
on tape.

Blackbox records each model call and tool call your agent makes. When a run goes wrong, rewind to the exact step, change one thing, and fly it again.

Bright orange so you can find it after a crash. Works with Anthropic and OpenAI message logs and OpenTelemetry traces.

PlaybackFlight: Elapsed 00:00.0Context 0 tokens

What goes on the tape

The same channels for every run, so you can line two flights up and see exactly where they split.

ChannelUnitWhy you want it
Model callsms, stop reasonWhat the model said, how long it thought, and whether it finished or hit a limit.
Tool callsname, input, outputEvery argument and every result, including the errors the agent quietly ignored.
Tokensin / outWhere the context ballooned, and which step made the run expensive.
CostUSD, estimatedPer step and per flight, at the prices you set.
Flagsloop, error, slow, claimRepeated identical calls, error streaks, slow tools, cut-off answers, and actions the model claims without a tool to back them.

Rewind. Change one thing. Fly it again.

Pick any step and fork. Edit the prompt, a tool result or the model, and Blackbox replays the run from there as a new branch you can compare side by side.

originalfork: tool result editedstep 6 loop x3passes
  • LiveBuilt-in tools like run_js and fetch_url run for real.
  • ReplayIf the forked run makes a call it already made, the recorded result comes back.
  • SimulatedFor anything else, Claude stands in for the tool using the recorded calls as reference. Simulated steps are clearly marked.

The incident report writes itself

Built-in detectors catch the obvious failures. Claude reads the whole tape for the rest and writes it up like an investigator.

FlightSupport bot claims a refund it never made
Steps9
SeverityHigh
Probable cause
After three identical lookup_order calls failed with a 503, the model told the customer a refund was issued. No refund tool was ever called.
Contributing factors
  1. Retries had no backoff and no change in input.
  2. No fallback message for an unavailable orders service.
  3. The system prompt forbids unissued refunds but nothing enforces it.
Recommendations
  1. Cap retries at two with backoff, then hand off to a human.
  2. Check every "refunded" claim against a successful issue_refund result before sending.

Bring your tapes

Anthropic messagesOpenAI chat + tool callsOpenTelemetry GenAI spansBlackbox JSON
Open the recorder