Skip to content
⚠️ Before You Start
Risk: Medium
Duration: Half a day workshop plus 1 day to document
Last Validated: Not yet validated (draft)
Kind: Decision guide
Risk Note: Medium (wrong metrics give false assurance to a client approving autonomous action)

Choose agent evaluation metrics, thresholds and judge models

Draft outline: not validated

Drafted with AI from official sources on 16 Oct 2026. Do not use on a live system until it is written, dry-run by someone other than the author and signed off (planned 2026-11-27). Author: Prasanna · dry-run: Devadharshini Raja · sign-off: Yogeshwaran. Work items: AGT-07-T09, AGT-07-T10, AGT-07-T11, AGT-07-T12.

1. Summary

At project start when agreeing acceptance criteria with a client, or when an agent moves up an autonomy level and needs stricter evidence.

Risk: Medium (wrong metrics give false assurance to a client approving autonomous action). Typical duration: Half a day workshop plus 1 day to document.

2. When to use and when not to

Use it when: At project start when agreeing acceptance criteria with a client, or when an agent moves up an autonomy level and needs stricter evidence.

Do not use it when: to be written by the author.

3. What you must know first

  • Outcome metrics (task success) and process metrics (tool selection, tool input accuracy, step efficiency) answer different questions
  • Binary pass/fail evaluators are easier to explain to clients than 1-5 scores
  • Deterministic checks (exact tool sequence) beat LLM judges where the right path is known
  • LLM judges need a stronger or different model than the agent under test
  • Cost per task needs token counts per run from traces

4. Options

Outline of the decision process; the author describes each option.

  1. List the decisions the client will make from the results, such as go-live or a higher autonomy level.
  2. Default to the four Level 2 metrics: task success, step efficiency, tool accuracy and cost per task.
  3. For each tool, decide between deterministic matching (strict, unordered, subset or superset) and an LLM judge.
  4. Pick the judge model and record its deployment and version.
  5. Set thresholds from a baseline run, not from guesses, and record them in the thresholds file.
  6. Add safety evaluation for agents that act on external content, using red-teaming results.
  7. Review the decision table with the QE lead and agent lead and sign it off.
  8. Re-run the decision when the agent gains a new high-risk tool or a new autonomy level.

Draft snippet: Level 2 metric and threshold record

{
  "agent": "invoice-triage-agent",
  "agentVersion": "1.4.0",
  "judge": {"deployment": "${JUDGE_DEPLOYMENT}", "isReasoningModel": true},
  "metrics": {
    "task_success": {"method": "llm_judge", "minPassRate": 0.9},
    "tool_accuracy": {"method": "llm_judge", "minPassRate": 0.95},
    "tool_sequence": {"method": "trajectory_match", "mode": "superset"},
    "step_efficiency": {"method": "custom", "minAverage": 0.8},
    "cost_per_task_usd": {"method": "trace_tokens", "maxAverage": 0.12}
  },
  "baselineRun": "eval/baseline_metrics.json",
  "approvedBy": ["QE lead", "Agent lead"],
  "reviewOn": "new high-risk tool or autonomy level change"
}

Snippets to add:

  • snippets/eval/thresholds.json: Metric and threshold decision record (json)
  • snippets/docs/eval-decision-table.md: Decision table: deterministic match versus LLM judge per tool (markdown)

5. Decision table

Situation Choose Why
to be written

To be written by the author from the outline above.

7. Known issues and fixes

Symptom Likely cause Fix
Client disputes a 1-5 score Likert scores are hard to defend Report binary pass rates per metric with example traces
Judge rates its own model's output too kindly Same model family judges itself Use a different or stronger judge and spot-check with human labels
Strict trajectory match fails valid alternative paths Wrong match mode for a flexible task Use unordered or superset match, or a trajectory LLM judge

8. Monitoring and alerts

Signal Alert when Action
Metric coverage report per release Any high-risk tool without a tool-accuracy test case Block the release until test cases exist

9. AWS and open-source equivalents

AgentCore Evaluations offers built-in evaluators plus custom judges on AWS. LangSmith agentevals provides the trajectory match modes; DeepEval separates trajectory metrics from component-level tool metrics in the same way.

10. Validation record

Pending. Planned dry-run scenario: A second engineer applies the guide to the sandbox demo agent and produces a thresholds file. Pass: every tool has a method, thresholds come from a recorded baseline run, and the QE lead can explain each choice from the decision table.

11. Sources