Evals
Score agent outputs, monitor quality, and route reviews.
Judge agent output quality. Create evaluators, monitor production traffic, and review uncertain results.
An evaluator uses a judge model to score each execution against its selected method. Configure the model, evaluation method, confidence threshold, and review routing in Control Center → Evals → Evaluators.
Judge model
The model that scores agent output
The model returns a rating and confidence
Evaluation method
Answer Relevancy, Faithfulness, Safety, or Task Completion
The method defines the parameters used to assess the output
Confidence threshold
A value from 0–100%. The default is 50%
Results below the threshold are flagged, regardless of rating
Human Review routing
Whether low-confidence results route to reviewers
A flagged result creates a Human Review task only when this setting is enabled
High-confidence results are accepted automatically. When routing is enabled, a reviewer records a verdict for each low-confidence result. The reviewer can pass the findings to the agent team to update regression tests, instructions, or knowledge.
Explore Evaluation
For common questions and fixes, see Evals: FAQs & Troubleshooting.
Related
Control CenterObserveControls & AuditAdminMaintained by Abhishek Paul
AI-assisted, human-approved
Last updated
Was this helpful?