Get Started with Judgment
Judgment helps teams improve production agents using evidence from real executions. When something goes wrong, it helps you understand why, measure how widely it occurs, and verify the change that addresses it.
Agent failures rarely live in the final response alone. They can begin with a missed tool call, stale context, faulty reasoning, or an execution path that only appears across a long trace. Judgment keeps the entire execution path available as evidence.
How Judgment works
Capture production executions
Trace the model calls, tools, state, and application logic that make up a complete agent run.
Investigate what went wrong
Use Judgment Agent to search across traces, find related failures, measure their impact, and narrow the root cause.
Define what matters
Turn the pattern into a judge with behavior outputs that you can evaluate consistently in production and offline.
Test and monitor the change
Preserve representative cases in a dataset, compare test runs, and watch the next release for regressions.
Quickstart
A working trace-and-evaluation loop, proven by a real trace and judge result.
Core workflows
Concepts
Learn how traces, judges, behaviors, datasets, tests, and monitoring fit together.
Production Monitoring
Detect recurring behaviors and get alerted when an agent regresses.
Investigate an issue
Move from one failed trace to a pattern, its impact, and a likely cause.
Test a change
Run production-derived cases against a candidate agent before release.