agent-exam

agent-exam

Eval framework for agent skills. Point it at a repo of skills and a suite of YAML tasks, and it runs them against a real agent CLI, grades the resulting transcript, and reports verdicts plus cost and token metrics.

Supported harnesses: Claude Code, Codex CLI, Copilot CLI and OpenCode. Each runs as a subprocess against your existing subscription or login — no pay-per-token API key is required.

Intended use

Interactive development of skills — edit a skill, run the relevant suite or a single task, read the report, iterate. Automating runs against subscription-backed CLIs may stretch a harness's terms, and LLM evals are slow and expensive enough that running every suite on every change buys little extra signal.

Install

pip install agent-exam

Quick start

In the repo holding your skills:

# evals/config.yaml
default_harness: claude_code

Skills are picked up from ./skills by default; set skills_dirs to point elsewhere.

# evals/suites/my-suite/tasks/greets.yaml
description: |
  The skill should greet by name.
kind: execute
prompt: |
  Greet Ada.
assertions:
  - skill_invoked: my-skill
  - judge: |
      The reply greets Ada by name.
  - no_permission_errors:

agent-exam doctor          # preflight checks
agent-exam my-suite        # run the suite
agent-exam show latest     # read the report

examples/config.yaml documents every configuration key, and the documentation covers writing suites, grading with judges and reading the reports.

The evals directory is found relative to the nearest pyproject.toml and defaults to evals/. Add a [tool.agent-exam] section only to point somewhere else, or to register a pre_run_hook:

# pyproject.toml
[tool.agent-exam]
evals_dir = "qa/evals"
pre_run_hook = "evals.hooks:pre_run_hook"

What it gives you

• Multi-harness by design. Assertions and the normalized transcript are harness-neutral, so adding another agent CLI is a small adapter rather than a core rewrite.
• Trigger evals are first-class. Test whether the right skill fires for a given prompt, separately from whether it then does the right thing.
• Reality-check modes. --without-skill re-runs a suite with one skill removed and --no-skills with the whole library removed, so you can see whether a skill earns its place against the bare agent.
• Tags for the costly evals. Tag a task or a whole suite, declare the tag excluded by default, and a run spanning every suite covers the cheap tasks only; --tag and --all-tags bring the rest back.
• Deterministic and LLM-judge assertions, per-attempt fixtures and isolated working directories, parallel execution with concurrency groups, and regression reports across runs (runs, show, history, diff).
• Rescoring without re-running. agent-exam rescore re-grades archived attempts against current assertions.

Contributing

uv sync
uv run pytest                  # unit tests
tox                            # full matrix, linting, packaging checks
pre-commit install

Pull requests run a secrets scan that must pass before merging; pre-commit install runs the same check locally. If it flags a value that is genuinely not a secret, audit it into .secrets.baseline and mention it in the PR.

License

Apache-2.0. See LICENSE.