Flight Recorder & Audit
Autopilot is built so you never have to take its word for anything. Every cycle is captured in the Flight Recorder: a complete, readable record of what each agent did, what it chose not to do, and the reason behind every decision. It’s a flight recorder, not a black box.
You’ll find it in the Harness panel, alongside an Activity timeline that spans every run.
How a run is recorded
- Run list — every cycle in order, each with a status dot (succeeded, running, or failed), a short summary, and a timestamp. While a run is in progress the view updates live.
- Run detail — open any cycle to see what triggered it (a scan, a schedule, or your manual instruction), the individual decisions, and the logs.
Decision cards — the heart of it
Each thing an agent decided becomes a decision card. This is where the “explains itself” promise lives:
| On the card | What it tells you |
|---|---|
| Outcome badge | Whether the decision was applied, was observe-only (proposed, nothing changed), or failed. |
| Action | What it did, in plain terms — “opened case: credential exposure,” “enabled secrets detector on wiki source.” |
| Rationale | A written reason for the decision. This is mandatory — every action and every deliberate non-action has one. |
| Details | The specifics behind the action, expandable when you want to dig in. |
Because non-actions are recorded too, you can see not just what Autopilot changed, but what it considered and deliberately left alone — and why.
Two ways to read the logs
Each run’s logs come in two channels, so the right person gets the right level of detail:
| Channel | Who it’s for | What it reads like |
|---|---|---|
| Business narrative | Analysts, reviewers, auditors | Plain English: “Reviewed 42 new findings, grouped 6 into an existing case, opened 1 new inquiry.” |
| Technical log | Operators debugging behaviour | A detailed, step-by-step trace of the run for when you need to see exactly what happened. |
Most of the time the business narrative is all you need. The technical channel is there for the moments you want to verify a specific step.
The run dashboard
The top of the Harness panel summarises Autopilot’s recent health at a glance:
| Metric | Meaning |
|---|---|
| Active runs | Cycles running right now |
| Runs (24h) | How busy Autopilot has been today |
| Applied | Changes actually made |
| Observe-only | Proposals that were logged but not applied |
| Failed | Decisions that couldn’t complete |
| Memory entries | How much Autopilot has learned |
| Brief version | How current the System Brief is |
A healthy instance running in managed mode shows mostly applied decisions; one you’ve kept in observe-only shows mostly observe-only — exactly the proposals waiting for your review.
Why this matters
- Trust through transparency. Nothing happens silently. If Autopilot opened a case or changed a source, the reason is one click away.
- Audit-ready. Every action is attributed and timestamped, so the record stands up to a compliance or security review.
- A feedback loop. Reading the rationale on a decision you disagree with tells you exactly which guidance or operator directive to adjust — see Steering & Fine-Tuning.
You’ve seen the whole picture
That completes the tour of Autopilot — from the big idea to the details:
| Page | |
|---|---|
| Overview | What Autopilot is and why it exists |
| Meet the Agents | The five agents and what each changes |
| How a Cycle Runs | Triggers, the run rhythm, and observe-only |
| Memory & System Brief | How it stays grounded and learns |
| Steering & Fine-Tuning | Every knob you can turn |
| Flight Recorder & Audit | How to read what it did (you are here) |