Evaluating AI Agent Quality with Agent Evaluation

    Who this is for:

    Architecture / Concept Overview: Evaluating AI Agent Quality with Agent Evaluation

    Agent Evaluation combines automated LLM-based judging with human review to produce actionable quality scores.

    %%{init: {"theme":"base","themeVariables":{"background":"#0B0E14","primaryTextColor":"#E0E6ED","lineColor":"#5D6470","darkMode":true,"primaryColor":"#2E4A4A","secondaryColor":"#374151","secondaryTextColor":"#E0E6ED","tertiaryColor":"#111827","tertiaryTextColor":"#E0E6ED","edgeLabelBackground":"#1f2937"}}}%% flowchart LR classDef source fill:#3F4B59,stroke:#9CA3AF,stroke-width:2px,rx:8,ry:8,color:#E0E6ED classDef ingestion fill:#5A4B36,stroke:#C9A86B,stroke-width:2px,rx:8,ry:8,color:#E0E6ED classDef processing fill:#535072,stroke:#8E82B4,stroke-width:2px,rx:8,ry:8,color:#E0E6ED classDef storage fill:#2E4A4A,stroke:#5FAFA8,stroke-width:2px,rx:8,ry:8,color:#E0E6ED classDef serving fill:#3D5550,stroke:#6BB7AA,stroke-width:2px,rx:8,ry:8,color:#E0E6ED classDef governance fill:#5A3F52,stroke:#C28BB0,stroke-width:2px,rx:8,ry:8,color:#E0E6ED AGENT[Deployed Agent] -->|Run| EVAL_SET[Evaluation Dataset] EVAL_SET -->|Generate| RESPONSES[Agent Responses] RESPONSES -->|Score| LLM_JUDGE[LLM-as-Judge] RESPONSES -->|Review| HUMAN[Human Review App] LLM_JUDGE -->|Metrics| REPORT[Quality Report] HUMAN -->|Feedback| REPORT REPORT -->|Gate| DEPLOY_GATE[Deployment Gate] REPORT -->|Monitor| PROD_MON[Production Monitoring] AGENT:::serving EVAL_SET:::source RESPONSES:::ingestion LLM_JUDGE:::processing HUMAN:::governance REPORT:::storage DEPLOY_GATE:::serving PROD_MON:::governance

    *Agent Evaluation pipeline: agents run against evaluation sets, LLM judges and humans score responses, and quality gates control deployment.*

    %%{init: {"theme":"base","themeVariables":{"background":"#0B0E14","primaryTextColor":"#E0E6ED","lineColor":"#5D6470","darkMode":true,"primaryColor":"#2E4A4A","secondaryColor":"#374151","secondaryTextColor":"#E0E6ED","tertiaryColor":"#111827","tertiaryTextColor":"#E0E6ED","edgeLabelBackground":"#1f2937"}}}%% graph TD classDef source fill:#3F4B59,stroke:#9CA3AF,stroke-width:2px,rx:8,ry:8,color:#E0E6ED classDef ingestion fill:#5A4B36,stroke:#C9A86B,stroke-width:2px,rx:8,ry:8,color:#E0E6ED classDef processing fill:#535072,stroke:#8E82B4,stroke-width:2px,rx:8,ry:8,color:#E0E6ED classDef storage fill:#2E4A4A,stroke:#5FAFA8,stroke-width:2px,rx:8,ry:8,color:#E0E6ED classDef serving fill:#3D5550,stroke:#6BB7AA,stroke-width:2px,rx:8,ry:8,color:#E0E6ED classDef governance fill:#5A3F52,stroke:#C28BB0,stroke-width:2px,rx:8,ry:8,color:#E0E6ED METRICS[Agent Quality Metrics] --> GROUND[Groundedness] METRICS --> REL[Relevance] METRICS --> SAFE[Safety] METRICS --> CORRECT[Correctness] METRICS --> COMPLETE[Task Completion] METRICS --> LATENCY_M[Latency] GROUND --> CONTEXT_SUPPORT[Response supported by context] REL --> ANSWER_QUESTION[Response addresses the question] SAFE --> NO_HARMFUL[No harmful content] CORRECT --> FACTUAL[Factually accurate] COMPLETE --> TASK_DONE[Task fully accomplished] METRICS:::governance GROUND:::processing REL:::processing SAFE:::storage CORRECT:::serving COMPLETE:::serving LATENCY_M:::ingestion CONTEXT_SUPPORT:::source ANSWER_QUESTION:::source NO_HARMFUL:::source FACTUAL:::source TASK_DONE:::source

    *Agent quality metric taxonomy: from groundedness to task completion.*

    Key Terms

    Prerequisites and Setup

    • databricks-agents package installed.
    • A deployed agent or model to evaluate.
    • An evaluation dataset with test questions and expected answers.
    • Foundation Model APIs for LLM-as-judge scoring.

    Step-by-Step Implementation

      Configuration Reference

      Evaluating AI Agent Quality with Agent Evaluation configuration options
      ParameterDefaultDescription
      model_nameUnity Catalog model to evaluate
      evaluation_setDataFrame or table with test cases
      metricsallSpecific metrics to compute
      enable_review_appfalseLaunch human review interface
      judge_modeldefaultLLM endpoint for automated judging

      Monitoring, Cost, and Security Considerations

      Common Pitfalls and Recommended Patterns

        Frequently Asked Questions