Outcome checks, trajectory review, sandboxes, simulated users and pass^k. A complete method for knowing whether your agent really works.
