Verification is the discipline of not taking the agent’s word for it: deterministic checks, real diff review, knowing what a green test suite does and does not prove, and catching the confident-but-wrong answer — the signature failure mode of capable models.
A green suite proves the code satisfies the tests that exist — nothing about behavior the tests never covered. After an agent refactor, the highest-value check is re-running the suite yourself and reading the diff for changes in behavior the tests do not constrain.
It looks like a clean explanation, a plausible diff, and a wrong result. Models present their best guess with the same fluency as their best knowledge. The tell is unverifiable specificity — exact numbers, file paths, or API names you did not give it and it did not read.
All of it — but with triage. Mechanical, repeated edits can be sampled once the pattern is proven; anything touching auth, money, data deletion, or public behavior gets read line by line. The expensive habit is skipping review because the agent “usually gets it right.”
A deterministic check gives the same verdict every run: a test, a type-checker, a curl returning 200, a byte-for-byte comparison. They convert “I think it’s fine” into evidence, and they let agents verify their own work mid-task instead of at the end.
An agent reports “all 42 tests pass” after a large refactor. What is the strongest next step?
B. B independently reproduces the claim and covers the gap green suites leave. A trusts an unverified report; C invites the same confidence that produced the claim; D adds volume, not verification.