# Evaluator-Optimizer Loop

> One model makes a draft and a second one judges it. They repeat until it passes.

**Field:** Artificial Intelligence  
**Category:** AI Agents  
**Also known as:** Evaluator-Optimizer, Generator-Critic Loop, Generator-Evaluator Pattern

## Think of it like

It is like a writer and a teacher. The writer hands in a page. The teacher marks it, and the writer tries again.

## Definition

An evaluator-optimizer loop is a pattern where one model makes something and another model judges it. The first model writes a draft. The second checks it against clear rules and sends back notes. The first model uses the notes to write a better draft. They repeat this until the judge is happy or a set number of rounds runs out.

## Example

Lucy's app translates poems. One model translates, a second model checks the tone and meaning and lists what is off, and after three rounds the judge says pass.

## Why it matters

Some work gets much better with honest notes. A separate judge with clear rules gives them every round.

## Common mistake

Giving the judge no clear rules. Vague notes like make it better send the loop round in circles.

## Three principles

- **Two roles.** One model makes, one model judges.
- **Clear rules.** The judge checks against stated goals.
- **A round limit.** Stop on a pass or after a set number.

## Related terms

- [Agent Reflection](https://learn.benibix.com/term/ai/agent-reflection)
- [LLM-as-a-Judge](https://learn.benibix.com/term/ai/llm-as-a-judge)
- [Multi-Agent System](https://learn.benibix.com/term/ai/multi-agent-system)
- [AI Workflow](https://learn.benibix.com/term/ai/ai-workflow)
- [Agent Stop Condition](https://learn.benibix.com/term/ai/agent-stop-condition)

## Told apart from

- [Agent Reflection vs Evaluator-Optimizer Loop](https://learn.benibix.com/compare/agent-reflection-vs-evaluator-optimizer): People think reflection and an evaluator-optimizer loop are the same thing, and they are not. In reflection, one model reads its own work and fixes it. In the loop, the job is split in two. One model writes, and a second model judges it by clear rules. A second judge can catch what the writer keeps missing.

![Three panels. Left: one job splits into three agents side by side, and their answers merge into one. Middle: a writer model passes a draft to a checker model, which sends notes back until it says pass. Right: one model writes a draft, reads it again, and fixes it.](https://learn.benibix.com/diagrams/ai7-agent-patterns.svg)

---

Source: https://learn.benibix.com/term/ai/evaluator-optimizer
Author: Behnia Zarei (a developer in Vienna, Austria, who builds and teaches data, AI and business intelligence at BENIBIX)
Publisher: BENIBIX Learn (https://learn.benibix.com/)
Last updated: 2026-11-19
Licence: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Quote freely with attribution.
