Eval

I tried to make an AI agent answer more. It answered less.

If you build or evaluate scoped agents: any talk about the agent in your test context makes it defend...