Agent Reliability Audit is a skill published by agensi that takes exported run logs, execution traces, tool calls, and retry records from AI agents and produces structured audit reports aimed at identifying failure modes before a production rollout.
The skill performs pattern detection across real-world transcripts to surface agent looping, behavioral drift, and latency hotspots that are not visible in unit test results. It also performs tool stability analysis by correlating a tool inventory against actual execution traces, flagging integrations that exhibit flaky or inconsistent behavior across multiple runs.
Audit output is generated in both Markdown and JSON formats. The Markdown report is structured for stakeholder review, while the JSON output is suitable for downstream processing or storage. Both formats include deep dives into recovery failures — cases where an agent failed to recover from an error state — and connect each observed failure pattern to a specific architectural improvement.
The skill is designed for Python-based workflows and is described as compatible with log output from LangChain, CrewAI, and custom OpenAI implementations. It is intended to fit into CI/CD pipelines as well as developer workstation workflows.
This skill is appropriate when preparing an AI agent for production and needing a systematic, multi-run reliability assessment that goes beyond ad-hoc log inspection. It is not a real-time monitoring tool and does not replace runtime observability infrastructure.