Review loop

A repeated evaluate-and-revise cycle that compares an agent result with explicit criteria, returns a specific failure with evidence, and stops or changes route under a defined rule.

A review loop begins with a checkable result. A writing workflow might detect that a recommendation is absent from the opening, attach the failed rubric item, revise that section, and rerun the check. The evaluator can be deterministic code, a source comparison, a calibrated model, a specialist, or a combination. The next pass improves only if the new evidence or instruction addresses the failure; repeated passes can also overfit or degrade a good result.

Builder example

A durable loop separates run repair from workflow maintenance. A one-off failure may need a revised output, while a recurring representative failure may justify a regression case, rule, tool change, or model change. Test proposed fixes against both the failed case and cases that already passed so a local correction does not create a wider regression.

Common confusion: Review is not synonymous with manual approval or with adding instructions forever. A schema validator can complete a review for structure, while factual and aesthetic properties need different evaluators. Every loop also needs a stopping, escalation, or cost rule.