Agents / Standard term
Review loop
A structured cycle where an agent runs, a person inspects the output, corrections feed back into the agent's instructions, and the next run improves.
A review loop is a structured cycle where an agent runs, a person inspects the output, corrections feed back into the agent's instructions, and the next run improves. After a writing agent drafts a report, the reviewer notes that it buried the recommendation in paragraph four. That correction becomes a permanent instruction: lead with the recommendation. Over successive cycles, the agent's instructions accumulate the reviewer's judgment and the error rate drops. The loop works only when corrections are specific enough to become rules and durable enough to persist across sessions.
Builder example
Teams that skip the review loop get stuck in a pattern: the agent makes the same mistakes, the person fixes them manually every time, and the automation never improves. In a review loop, you write each correction into the agent's instructions, so your judgment carries forward and the agent keeps that capability across runs. Without it, you are paying for AI-generated first drafts and doing the real work yourself.
Common confusion: A review loop is more structured than ad-hoc quality checks. It includes a defined review cadence, a format for capturing corrections, and a process for updating the agent's instructions or context files. Glancing at the output and thinking "looks fine" is not a review loop.