Judgment documentation
Judgment is infrastructure for engineering long-horizon agents.
Set up Judgment with your coding agent
Paste it into your agent from your repo's root.
Judgment works from your agent's traces, whether they come from production or from tests in your own environment.
Products
Tracing
Send your agent's traces to Judgment from Python or TypeScript.
Issues
Find recurring problems in your agent's traces that no judge is checking for yet.
Judges
Label traces and sessions with labels you define.
Judgment Agent
Ask about your traces and labels in plain words.
Tests
Run offline tests on your agent with a test suite, and check every run with your judges.
MCP server
Connect your coding agent to Judgment over MCP.
Start here
Quickstart
Set up your first project and see what Judgment finds in your agent's traces.
Example agent
The lunch-order agent that the docs use as a running example.
How it works
Discover with Issues
Agents search your traces for recurring problems you didn't know to look for.
Monitor with judges
A judge labels traces for a failure or signal you already know to look for, and can alert your team when it fires. Most judges are Agent judges, so they can evaluate long traces.
Analyze with the Judgment Agent
Judge labels act as indexes. The agent uses them to answer questions across all your data and cites the traces behind each answer.
Test before you ship
Tests run your agent over a set of test inputs and grade each trace with the same judges.
Confirm the fix
The same judges keep labeling new traces, so you can see whether the failure stops.