A long delegated change is finished and four ways of checking it are proposed. Which one puts something other than the agent that did the work in the grading seat?

The principle — moving the grading off the worker is a claim about two things at once: who is doing the grading, and what they are grading. Get either wrong and the check is decorative.

Why the key is correct — Anthropic describes the design in one sentence: a verification subagent or a dynamic workflow that checks its own findings has a fresh model try to refute the result, so the agent doing the work is not the one grading it. Two properties are packed in there. The grader is a fresh model rather than the worker, and what it is handed is the result — so an attempt to refute has something falsifiable to work against. The glossary supplies the reason it matters: without verification, the only thing deciding the agent is finished is the agent itself. The same sentence offers a second shape of the same design, a dynamic workflow that checks its own findings, so this is a pattern rather than a single feature.

Why the others are wrong — a critical re-read keeps the author in the seat, and a fresh context improves review precisely because Claude will not be biased toward code it just wrote. Confirming against its own record has the same defect with more paperwork; the guidance asks Claude to show evidence rather than assert success — the test output, the command it ran and what it returned. The second subagent reading the report is the near miss worth studying: it does introduce a different grader, but it grades a summary. A summary is what the first agent says about the work, so a check against it can only find the summary self-consistent. Refutation needs the artefact, not the account of it.

Remember this — an independent check has to be someone else, looking at the work — not someone else looking at the story about the work.

Sources — Anthropic's Claude Code best practices and glossary.