# Agent Evaluation

Source: https://www.potatoannotator.com/docs/agent-evaluation/overview

This section treats a model's output as something to be reviewed rather than something to be labelled. The unit of work is usually a trace: one run of an agent, with its tool calls, intermediate steps and final answer.

[LLM-as-Judge Calibration](/docs/agent-evaluation/judge-calibration) and [Judge ↔ Human Alignment](/docs/agent-evaluation/judge-alignment) cover the case where a model does the grading and you need to know how far to trust it. [Three-Pane Trace Evaluation](/docs/agent-evaluation/eval-trace) covers the case where a person does. [CI Evaluation](/docs/agent-evaluation/ci-evaluation) covers running either on every commit.

Potato does not train models. Everything here produces labels, scores and agreement statistics; what you do with them afterwards happens outside the tool.