| name | eval-agent |
| description | Build LLM-as-judge components from real references: gather examples, generate rubrics, score outputs, return findings, and wire improvement loops. |
Eval Agent
Use this when the user needs a judge component, rubric, scoring loop, or evaluation report for subjective outputs.
Do not use LLM judges for objectively checkable behavior; write tests instead.
Preconditions
- The target is subjective enough that two humans could reasonably disagree.
- Real reference material exists or can be gathered first.
- The evaluator output will drive a decision, not decorate a dashboard.
Build Flow
- Gather positive examples, negative examples, domain docs, and known failure cases.
- Generate rubric dimensions from those references; do not hand-write vibes.
- Define the target schema, finding schema, score range, and pass/fail threshold.
- Run a small calibration set and inspect disagreements before broad scoring.
- Wire the evaluator into the smallest loop that will improve the target.
- Persist prompts, references, scores, findings, and model snapshot.
Rules
- Quote or cite the references that justify each rubric dimension.
- Keep judge prompts deterministic and versioned.
- Separate single-output scoring, per-turn scoring, and reviewer-driver loops.
- Report disagreement, uncertainty, and failure modes; do not hide low-confidence scores.
- Replay or re-score from saved artifacts when possible instead of spending live model calls.
Use references/full-reference.md for component examples, prompt patterns, and the old full guide.
Then consider
agent-eval when the task touches the eval package internals.
evolve when the evaluator is ready to optimize a measurable target.