Evals are tests. Treat them like it.

An eval that nobody runs is a document. An eval that runs on every pull request is a test, and it changes how a team talks about quality.

Keep cases small

One behaviour per case. When a case fails you should know what broke from its name alone.

Review them like code

Cases live next to the agent, change through pull requests and carry an owner. Deleting a case needs a reason in the commit message.

Let them block

Set a pass-rate gate and make it required. The first week is noisy. The second week is the first time anyone trusts a green check.