Evals work best when they enforce consistency in exactly the places where consistency actually matters — not everywhere an agent acts. Linear's framing: 'evals are most successful when you're ensuring consistency somewhere that consistency is important. But consistency is not always important for agents.' Agent behavior has legitimate variance (tone, phrasing, exact steps taken) that shouldn't be pinned down by an eval; over-evaluating that variance produces false failure signals and pressure to over-constrain the agent. The practical move is to ask, before writing an eval, whether the target behavior actually needs to be consistent — and if not, skip it. Complements Code-Based (Reference-Based) Evaluation, LLM-as-Judge Validation, and Eval Investment Triage.