No Agent Reviews Its Own Work

Описание и применение

Внедрение четырёх ролей ИИ (исполнитель, рецензент, тестировщик, визуальный регресс) и человеческого контроля обеспечило безопасную миграцию Angular без простоя, повысив трассируемость и надёжность изменений.

Задача

Migrating a critical enterprise healthcare platform from Angular 11 to 15 without downtime while ensuring AI-generated fixes are correct and traceable.

Результат

Implemented a four‑role AI workflow that reduced manual effort, increased traceability, and prevented silent regressions during a critical, zero‑downtime Angular migration.

Рабочий процесс

• 1. Executor agent upgrades Angular and resolves build errors.
• 2. Reviewer agent examines executor’s changes for correctness, convention adherence, and root‑cause resolution.
• 3. Test agent runs the relevant verification suite (unit, integration, smoke).
• 4. Visual regression agent compares UI renders across versions and reports differences.
• 5. Human reviewer assesses all outputs, applies a checklist, and decides whether to advance to the next version.

Доказательства

The hard part is different: how do you know the agent's fixes were correct?

Источник

In one migration step, the reviewer flagged aspects of the executor's changes and gave reasons for its objections.

Источник