# Orchestrator

URL: https://codepanel.ai/glossary/orchestrator

In a panel review, the model that merges every reviewer's findings, drops what the code contradicts and writes the one review that is posted.

Reviewers are tuned to report everything they notice; the orchestrator is tuned to keep only what the code supports. It sees the reviewers' findings and the code around the files they name, never the whole repository, and it also rules on any finding the coding agent disputes.

## In an agent review loop

Three reviewers read a pull request in parallel. Two report the same unchecked `null` in different words, and the third reports a SQL injection in a query that is already parameterised. The orchestrator is shown the diff and the code around the files they name. It keeps one copy of the `null` finding, as F1, and drops the injection claim because the code it was shown contradicts it. The coding agent sees one finding, not three.

Later the agent disputes F1, citing a guard two lines up. About three minutes after the dispute, the orchestrator reads the finding, the rebuttal and the code at that commit, and withdraws, revises or upholds it. CodePanel prefers an orchestrator from a different model family than the reviewers that raised the finding, so it is not judging its own work.

## Where it fits

The orchestrator only sees the files the findings name, so its cost grows with what the reviewers report, not with the size of the change. [How it works](/docs/how-it-works#the-orchestrator) describes its inputs, and [one model or a panel](/guides/single-model-vs-panel) weighs the design against a single reviewer.