Agent Eval Coverage Audit

Audit your AI agent's evaluation coverage to identify missing release gates and production risks.

Install
cmdop skills install agensi-agent-eval-coverage-audit

Agent Eval Coverage Audit is a skill that inspects an AI agent’s testing infrastructure and produces a structured remediation plan. It examines evaluation configurations, sample datasets, CI/CD hooks, and policy checks to locate gaps that could allow a flawed agent build to pass release gates undetected.

The skill works with any JSON-based evaluation configuration, including formats used by tools such as Promptfoo and LangSmith. It cross-references an agent’s defined success criteria against existing traces and configurations to surface “false greens” — tests that appear to pass but do not catch real failure modes — and missing edge cases that represent production risk.

Once the audit completes, it generates two outputs: a Markdown report formatted for sharing with stakeholders, and a machine-readable JSON file suitable for integration into CI/CD pipelines. The supported runtime environments are PowerShell and Python 3.x.

This skill is appropriate when a team is preparing an agent for production and wants an automated, repeatable check of their eval suite rather than relying on manual review. It is not a runtime monitoring tool and does not execute the agent under test itself; it analyzes the evaluation infrastructure around the agent. Teams that have no existing eval configuration or traces will have limited material for the audit to inspect.

Use cases

  • Audit an eval suite before promoting an AI agent from pilot to production
  • Identify missing edge-case tests in an existing JSON-based evaluation configuration
  • Detect false-green test results that mask real failure modes
  • Generate a Markdown gap report for engineering and product stakeholders
  • Produce a machine-readable JSON remediation plan for automated CI/CD gating
  • Validate that policy checks and CI/CD hooks are wired into the release process

When to use it

  • The agent is approaching a production release and the eval suite needs a structured review
  • The team suspects their current tests pass too easily without catching real failure modes
  • A CI/CD-integrated quality gate is needed for agent releases
  • The evaluation configuration is JSON-based and compatible with Promptfoo, LangSmith, or similar frameworks

When not to use it

  • No evaluation configuration or trace data exists yet — the audit needs existing material to inspect
  • Runtime monitoring of a live agent is required rather than pre-release eval review
  • The agent's eval configs are not JSON-based and cannot be parsed by this skill
  • The environment does not support PowerShell or Python 3.x