Agent Trajectory Failure Forensics

How the workflow runs

1. 1Normalize heterogeneous logs into a stable ordered trajectory without changing speaker or tool semantics.
2. 2Derive static task, policy, tool, and structural invariants.
3. 3Evaluate context-sensitive invariants at every step and record evidence for violations.
4. 4Separate candidate failures from consequences and localize the earliest critical step.
5. 5Classify the failure with a grounded taxonomy and calibrated confidence.
6. 6Produce a replayable report, targeted regression case, and remediation priority.

Example usage

Start with one of these requests, then replace the details with your own.