# Agent run

The plan's exit test for agents, on the real CLI: a fresh agent gets only
`tasks.md` and `../../GUIDE.md` and does the tasks in a sandbox.

1. `tests/agent/setup.sh <sandbox>` builds the sandbox and leaves its server running. It prints the host; `<sandbox>/facts` has the values `tasks.md` names in braces.
2. Give a fresh agent `tasks.md` with `{SB}`, `{H}`, `{BANNER}` and `{DRAFT}` filled in, followed by the guide.
3. `tests/agent/grade.sh <sandbox>` checks the end state and stops the server. Read the agent's report for what it told the user and which flags it used.

Results are in `../dry-run-results.md`.