Agent Black Box
Look Past
“Everything Worked.”
Follow recorded model calls, tool activity, policy decisions and outcomes. Distinguish what the platform observed from what another producer reported.
Controlled Recorder Fixture
What Is Happening Here?
The example is deliberately small. These are the parts to understand before bringing your own task.
Reconstruct The Sequence
Inspect recorded attempts, failures, approvals and the available effect evidence in one run history.
Know The Basis
Observed, reported and inferred are different evidence categories. A completed tool call is not automatically a verified external effect.
Inspect Beyond The UI
Export evidence bundles and check hash chains and signed checkpoints using the supported verification tools.
Where It Stops
Recording and capture settings determine what exists. Redaction, sampling and retention can leave gaps. Integrity checks do not prove correctness or completeness, and Black Box does not itself authorize or prevent actions.