agent-introspection-debugging

Agent introspection works by capturing the failure state of an agent run with precise error details and context, analyzing the captured data to diagnose root causes, and applying contained recovery to resolve repeated failures or prompt drift.

The best way to diagnose root causes of repeated agent failures is to capture the failure state and analyze the data using a structured template to identify root causes, apply contained recovery, and generate a human-readable introspection report for escalation.

You generate an introspection report for agent failures by describing the failure context and running the introspection workflow, which specifies minimum data to capture and produces a human-readable report for escalation.

You should use an agent introspection workflow for debugging when tool calls loop, prompts drift, or environment changes cause degradation, providing safe escalation and traceability through structured diagnosis and recovery.

No specific dependencies are required to run agent diagnosis and recovery. The workflow uses a structured template for diagnosis and recovery steps to capture failure data and produce a human-readable debug report.