| name | agent-eval-design |
| description | Design rigorous evaluations and benchmarks for AI agents, developer tools, retrieval systems, and repository-scale automation. Covers task selection, contamination control, metric choice tied to engineering decisions, and statistical validity. Use when asked to design an eval/benchmark, critique an existing benchmark, choose metrics for an agent or RAG system, or decide whether a measured improvement is real. NOT for running an existing performance-benchmark suite or a per-feature acceptance checklist, or one-off model spot-checks. |
| scope | universal |
| ported-from | global skill |
Agent Evaluation & Benchmark Design
Design evaluations for AI agents, dev tools, retrieval, and repo-scale
automation. Evaluation quality beats benchmark size — a small set of real,
uncontaminated, decision-driving tasks is worth more than thousands of synthetic
puzzles.
Core principles
- Measure real tasks. Representative workloads over synthetic puzzles. If no
one does the task in practice, the number is noise.
- Separate capability from prompt engineering. Hold the harness/prompt fixed
when comparing models; hold the model fixed when comparing prompts. Report which
you varied. A gain you can't attribute is a gain you can't ship.
- Every metric must change an engineering decision. Before adding a metric,
name the decision it informs. If nothing changes based on its value, cut it.
- Reproducibility is a first-class result. Pin model versions, seeds, dataset
hashes, harness commit, and date. An unrepeatable eval is an anecdote.
- Minimize contamination. Assume public benchmarks are in training data;
prefer held-out, private, or post-cutoff tasks and say so.
Dimensions to consider
Correctness · completeness · reliability · latency · cost · determinism ·
reproducibility · developer effort · failure recovery · robustness. Pick the few
that map to real decisions for this system; don't report all ten by reflex.
Repository-scale evaluations
For agents that operate over codebases, evaluate the axes that synthetic tasks
miss: repository understanding, cross-file reasoning, architectural consistency,
migration quality, semantic correctness (not just diff-match), dependency
propagation, test generation, and documentation accuracy. Verify outcomes by
execution (tests pass, build green, behavior preserved) rather than string
similarity to a reference solution.
Benchmark design — audit before trusting
Before believing a benchmark, check it for:
- Contamination / dataset leakage — is the answer reachable from training
data or from the prompt itself?
- Unrealistic tasks — puzzle-shaped work no engineer actually does.
- Missing edge cases — the failure modes that matter live in the tail.
- Insufficient statistical power — enough trials and items to distinguish
signal from run-to-run variance? Report variance/CIs, not a single point.
- Evaluation blind spots — what the metric structurally cannot see (e.g.
pass@1 hides flakiness; exact-match hides correct-but-different solutions).
Metrics: engineering value, not leaderboard rank
Prefer metrics that capture value delivered: task completion, semantic
correctness, regression rate, benchmark coverage, engineering effort saved,
implementation quality, verification quality, cost-performance tradeoff.
A metric that only moves a leaderboard and no product decision is a distraction.
Guard against parity-before-efficiency errors: confirm two systems produce
equivalent outputs on matched inputs before comparing their speed or cost.
A cheaper system that quietly drops work is not cheaper.
Tracking across runs
Role: read-only analysis/planning — route implementation of the eval harness to a
separate pass. Store the pinned baseline (dataset hash, harness commit, model
version, seeds, date) somewhere durable in the repo (e.g. evals/baseline.json)
and have every recurring run diff against it. Emit metrics as a tracked series,
not a one-shot number — "is the improvement real" is answered against history.
Output
Recommend, with evidence:
- Improved benchmark design — concrete changes to tasks, splits, or scoring.
- New evaluation methodologies — where the current approach is structurally blind.
- Missing metrics — each tied to the decision it would inform.
- Stronger validation — contamination controls, statistical power, repro pinning.
- Cost-performance framing — the tradeoff curve, not a single winner.
- Opportunities — publication or product angles the eval surfaces, when they exist.