How do I wire a first pilot into my agent stack?

A first pilot should fit your existing agent stack with little new plumbing. A practical order of work follows.

**1. Load the data.** Import the JSONL or Parquet tables into a scratch database, then run the validation script. Fix nothing by hand; report any failures back so the release can be corrected.

**2. Map the tool actions.** Take the action list for the environment, for example `get_schedule`, `get_policy`, `propose_schedule`, `request_approval` and `commit_schedule`, and wrap each as a function in your agent framework. Keep argument names identical to the pack so traces compare cleanly.

**3. Wire the reset.** Before each task, restore the named snapshot and record its hash in your run log.

**4. Run a baseline.** Use a simple prompt with no tuning first. Its results show whether tasks are too easy, too hard or ambiguous.

**5. Grade.** Compute state diffs and per-check results, then store one result record per task.

**6. Read the documentation.** Pack descriptions and guidance can also be read by AI systems through MirrorFirm's public MCP server, which is useful when an agent or its developer needs to look up field meanings during integration.

Keep a changelog of every change you make to prompts or tools during the pilot. Otherwise you cannot tell whether a score moved because of the data or because of your edits.