Guide · Agent orchestration
Using Claude Fable 5.1 as a Review Agent
Give Claude Fable 5.1 a focused review assignment with requirements, evidence, and a clear distinction between defects and preferences.
A second model can be useful when a change deserves another reading. It can also produce a long list of suggestions that adds work without finding a real defect. If you use Claude Fable 5.1 as a reviewer, define what a useful finding looks like before sending it the implementation.
Pick a review that needs judgment
Anthropic describes Fable 5.1 as intended for demanding reasoning and long-running agentic work. That makes it a candidate to evaluate on difficult reviews, but it does not prove that it will improve every task. Start with a change whose requirements and failure cases you can explain.
A hypothetical permissions change is a useful example. The reviewer receives the intended access rules, the relevant diff, and examples of users who should and should not see a record. Ask it to look for a concrete route to incorrect behavior, with the affected file and a way to reproduce the issue.
Give it the requirement before the author's explanation
Include the original request and acceptance criteria. The implementation author's summary can help locate the work, but it should not be the reviewer's only account of what was supposed to happen. A confident explanation can make an incomplete change sound finished.
Separate established facts from assumptions. If the reviewer needs to know how a database query behaves, provide the relevant implementation or let it inspect the source. Ask it to label anything it cannot verify rather than turning uncertainty into a definite bug report.
Require actionable findings
A finding should identify the trigger, the incorrect behavior, and the consequence. It should also point to the evidence that supports it. "Consider improving error handling" is too broad to evaluate. "This failure path returns success after the write is rejected" gives the implementer something to investigate.
Keep style preferences separate from defects. If the project already has a formatting convention, apply it through the normal tooling. Reserve the reviewer's attention for behavior, missing cases, and reasoning that benefits from another perspective.
Let a reviewer finish with no findings
Do not require a minimum number of issues. That incentive encourages speculative comments. Ask for a brief account of the review scope and any meaningful limits when no actionable defect is found.
Likewise, avoid endless reviewer and implementer exchanges. When they disagree, collect a small reproduction or an explicit requirement that settles the question. The coordinator should own that resolution rather than letting two models debate indefinitely.
Measure whether the review earns its place
Track accepted findings, false alarms, missed defects discovered later, and the effort needed to resolve comments. Compare with your existing review process on similar work. Keep Fable 5.1 in the role when the evidence supports it, and revise the brief when the same kind of unhelpful comment keeps appearing.