AI Agent Failure Forensics

AI Agent Failure Forensics is an evidence-first diagnostic Skill for investigating why AI agents fail, behave inconsistently, produce incorrect results, misuse tools, or break during multi-step workflows.

Instead of simply asking an AI to “debug the problem,” it reconstructs the execution sequence, compares expected versus observed behavior, identifies the first meaningful divergence, separates symptoms from root causes, evaluates competing hypotheses, and produces an evidence-supported diagnosis.

The Skill is designed for developers, AI engineers, agent builders, QA teams, and technical teams working with agentic systems.

It delivers a structured Failure Evidence Pack containing the failure classification, execution reconstruction, first divergence, evidence chain, root cause, failure cascade, recommended remediation, regression test, verification status, and residual risk.

The workflow also includes explicit safeguards for untrusted logs and artifacts, prompt-injection resistance, evidence integrity, uncertainty handling, least-privilege analysis, privacy protection, and verification discipline.

Use it when you need to answer:

“What actually went wrong, where did the agent first diverge from the expected path, why did it happen, and how do we prove the fix works?”

This is not a generic debugging chatbot or prompt generator. It is a reusable forensic workflow designed to turn ambiguous agent failures into defensible, actionable, testable, and verifiable diagnoses.