For people building AI agents: what is the smallest evaluation you run before trusting an agent with a real task? I’m especially interested in how you test failure recovery and unsafe tool calls.