On this page
Learning objectives
- Build task-specific offline and online evaluations
- Trace model, tool, state, policy, and human decisions
- Turn incidents and reviewer corrections into regression cases
Prerequisites
A versioned agent spec · Representative task examples · A grading rubric
Answer first
Canonical definition
Agent evaluation measures whether a system completes a defined task within quality, policy, cost, and reliability bounds; observability records the run so failures can be diagnosed and reproduced.
Operating context
Why it matters
A final answer score cannot explain whether the agent retrieved bad evidence, selected the wrong tool, repeated an action, or exceeded its mandate.
Production traces become useful only when they are tied to versions, user outcomes, and a failure taxonomy.
Method
Step by step
- 01
Build the case set
Include common tasks, hard cases, policy boundaries, missing data, tool failures, malicious inputs, and costly mistakes.
- 02
Grade multiple dimensions
Score task result, evidence, policy compliance, tool choice, state change, latency, and full cost.
- 03
Trace the run
Correlate model requests, retrieved context, tool calls, validations, approvals, state transitions, and final outcome.
- 04
Close the regression loop
Convert production incidents, rejects, and surprising successes into versioned tests before the next release.
Hands-on lab
Create a 25-case agent eval
Assemble a balanced test set for one agent and compare two prompt or model versions.
Deliverable
A rubric, labeled cases, automated checks, trace links, score report, and release threshold.
What breaks
Common failure modes
- F1Evaluating only the final prose and ignoring actions
- F2Using a model judge without calibration against human decisions
- F3Collecting traces that omit versions, permissions, or state changes
Beyond the demo
Production notes
- Redact or hash sensitive data while keeping enough identifiers to investigate authorized incidents.
- Monitor distribution shift, tool error rates, human override, cost, and stuck-loop signals—not only average score.
Further reading
Sources
Related Tenten resources