What we learn testing AI agents: how we build tasks, what our evals catch, and the writes that go wrong.
The agent is non-deterministic. The consequence is not.