session-archaeology

To extract execution patterns from agent session transcripts, parse tool calls, model responses, and errors into structured event streams to discover recurring sequences. This process applies cross-session correlation with minimum frequency thresholds to identify high-confidence workflows.

Failure-mode taxonomy in agent session analysis categorizes recurring mistakes such as hallucination, tool misuse, and context loss. It attaches suggested fixes to each categorized error to enable faster remediation and prevent future regression in production agent workflows.

Yes, you can detect regression in production AI agents by mining JSONL logs to track behavior trend changes over time. Comparing recent session events against historical patterns reveals when previously successful workflows degrade or fail.

Session analysis requires an anonymization-first workflow that strips sensitive information from transcripts before cross-team analysis or sharing. This reduces risk while allowing teams to collaboratively review tool call traces, errors, and outcomes.

The best way to codify high-frequency agent patterns into candidate skills is to apply minimum frequency thresholds to cross-session correlations. This ensures only high-confidence, recurring execution sequences are transformed into proposed skill updates.