aurat.Open workspace Replay the run.
Catch the regression.
Your agent changed.
Know exactly what broke before you merge.
01 โ BEHAVIOR, REPEATED.A number is not a summary.
The model returned summary: 42. Scout accepted it.
Aurat caught the invalid outcome.
{ "rejected": true }{ "rejected": false }The answer can look right.
The workflow can still be wrong.
A ticket created twice.
A required step skipped.
A tool given the wrong argument.
Aurat tests the application around the model. Capture a known workflow, define its expected behavior, and make the next change with evidence.
From a real run
to a reliable test.
Keep your model SDK and tracing stack.
Add a repeatable check where you ship.

aurat proxy --mode record // Model traffic + wrapped tools // Review fixtures before committing.
A real code defect.
Four unchanged fixtures.
We ran Aurat against Scout's existing narration module. Three malformed responses exposed missing type validation.

Same requests. Same recorded responses. The actual Scout function executes in both cases.
Read the reproduction Source-discovered defect with synthetic model responses. Not a claimed customer production incident.Know what changed.
And why it matters.
Open a run, inspect the failed invariant, and compare it with a passing baseline. Model and tool coverage make missing interactions visible.
- Readable assertion failures
- Scenario-level evidence and events
- Versioned reports from your own CI


A small addition
to the way you work.
Start locally. Commit your scenarios. Run them in GitHub Actions, then bring the report into your workspace.
Read the integration guide