evaluation

Agent performance measurement uses multi-dimensional rubrics that score factual accuracy, completeness, citation accuracy, source quality, and tool efficiency. Define these dimensions in a rubric, create a test set with expected outputs, and run the evaluation runner to generate per-dimension scores and an overall weighted score.

Continuous production monitoring tracks agent outputs against your rubric dimensions over time, revealing regressions and configuration drift. Run evaluation across different tasks and configurations to detect when context changes degrade performance and maintain deployment readiness.

Build a test set with input prompts and expected outputs, define weighted scores for each evaluation dimension, then execute the evaluation runner. The runner generates per-dimension results and an overall score, enabling systematic comparison across agent versions and configurations.

Yes. Continuous production monitoring stores per-dimension and overall scores across deployments, letting you detect performance trends, identify regressions in factual accuracy or tool efficiency, and make data-driven decisions about configuration changes or model updates.

The standard evaluation dimensions are factual accuracy (correctness of information), completeness (coverage of the task), citation accuracy (proper source attribution), source quality (reliability of references), and tool efficiency (effective resource and API usage). Weight each dimension based on your task requirements.