Agent Benchmark Framework
Without benchmarks, we cannot know whether agent changes improve or degrade quality. This skill defines how to measure, track, and protect agent performance.
When to Activate
- Before and after modifying any agent definition file
- When adding a new skill that an agent depends on
- Periodic quality audits (weekly/monthly)
- When a user reports degraded agent output
- Before promoting an agent from experimental to production
Core Concepts
Why Benchmarks Matter
Agent quality degrades silently. A prompt tweak that improves one response can break ten others. Without a baseline to compare against, every change is a guess. Benchmarks make quality visible and regressions detectable.
Benchmark Types
| Type | Scope | Cost | Frequency |
|---|---|---|---|
| Prompt Benchmark | Single agent, single task | Low | Every agent change |
| Task Benchmark | End-to-end scenario | Medium | Feature changes |
| Regression Suite | All critical agents | High | Weekly / before release |
Directory Structure
~/.claude/benchmarks/ fixtures/ code-reviewer/ missing-error-handling.ts # Input: code with no try/catch sql-injection.py # Input: unparameterized query clean-code.ts # Input: code with no issues security-reviewer/ hardcoded-secret.ts # Input: API key in source parameterized-query.py # Input: safe query (no findings expected) verifier/ passing-build/ # Input: project that builds failing-types/ # Input: project with type errors ground-truth/ code-reviewer/ missing-error-handling.json # Expected findings sql-injection.json # Expected findings clean-code.json # Expected: empty findings security-reviewer/ hardcoded-secret.json parameterized-query.json rubrics/ code-reviewer.md # Scoring rubric security-reviewer.md verifier.md baselines/ code-reviewer-2026-03-01.json # Timestamped baseline scores code-reviewer-2026-03-26.json security-reviewer-2026-03-26.json results/ run-2026-03-26T14-00.json # Latest run output
Scoring Rubric Template
Each agent has its own rubric file. The template:
markdown## [Agent Name] Scoring Rubric ### Completeness (0-30 points) Did the agent find everything it should have found? - Found all expected issues: 30 - Missed 1 non-critical issue: 22 - Missed 1 critical issue: 10 - Missed 2+ issues: 5 - Found nothing when issues exist: 0 ### Accuracy (0-30 points) Were the findings correct? No false positives? - All findings verified correct: 30 - 1 false positive: 22 - 2 false positives: 12 - 3+ false positives: 5 - Majority of findings are wrong: 0 ### Actionability (0-20 points) Did the agent give concrete, implementable fixes? - Clear fix with file/line reference: 20 - Clear fix without location: 14 - Vague suggestion (refactor this): 7 - No fix suggested: 0 ### Format Compliance (0-20 points) Did the output follow the agent's output contract? - Matches contract exactly (VERDICT + sections): 20 - Minor deviation (missing one section): 12 - Major deviation (no VERDICT): 5 - Unstructured free text: 0
Ground Truth Format
Ground truth files define what a correct agent response must contain:
json{ "fixture": "missing-error-handling.ts", "agent": "code-reviewer", "required_findings": [ { "id": "missing-try-catch", "severity": "HIGH", "description_contains": ["error handling", "try", "catch"], "location_hint": "fetchUserData" } ], "forbidden_findings": [], "required_verdict": "FAIL", "min_score": 70 }
Scoring Logic
How a Run Is Scored
1. Load fixture (input code / task) 2. Run agent with fixture as input 3. Parse agent output 4. Check required_findings: each found = +completeness points 5. Check forbidden_findings: each false positive = -accuracy points 6. Check verdict matches required_verdict 7. Check format follows output contract 8. Sum scores → final 0-100 9. Compare against min_score threshold
Score Interpretation
| Score | Status | Action |
|---|---|---|
| 90-100 | EXCELLENT | No action needed |
| 75-89 | GOOD | Minor tuning optional |
| 60-74 | WARN | Investigate degradation |
| 40-59 | POOR | Agent needs rework |
| 0-39 | CRITICAL | Block deployment |
Running Benchmarks
Run All Benchmarks
bash# Full suite node ~/.claude/benchmarks/run.mjs # Output: results/run-{timestamp}.json
Run Single Agent
bash# Benchmark one agent node ~/.claude/benchmarks/run.mjs --agent code-reviewer # With verbose output (shows actual vs expected per fixture) node ~/.claude/benchmarks/run.mjs --agent code-reviewer --verbose
Compare Against Baseline
bash# Compare latest run against saved baseline node ~/.claude/benchmarks/run.mjs --compare # Compare specific run against specific baseline node ~/.claude/benchmarks/run.mjs \ --compare results/run-2026-03-26.json \ --baseline baselines/code-reviewer-2026-03-01.json
Update Baseline
Only run this after verifying an improvement is real:
bash# Promote latest results to new baseline node ~/.claude/benchmarks/run.mjs --baseline update # Creates: baselines/{agent}-{date}.json
Regression Detection Rules
A regression is triggered when:
- Score drops more than 10 points on any single fixture
- Average score drops more than 5 points across all fixtures for an agent
- A previously PASS fixture becomes FAIL
- Format compliance drops below 80 (agent stopped following output contract)
Regression Report Format
REGRESSION DETECTED: code-reviewer Fixture: sql-injection.py Baseline score: 88 Current score: 61 Delta: -27 (CRITICAL) Missing finding: SQL injection in execute_query() line 14 Root cause: Agent definition changed, removed security focus Recommendation: Revert agent change or add SQL injection examples
Metrics Tracked Per Agent
| Metric | Formula | Target |
|---|---|---|
| accuracy | correct_findings / total_findings | >= 0.85 |
| completeness | found_issues / total_issues | >= 0.90 |
| false_positive_rate | false_positives / total_findings | <= 0.10 |
| format_compliance | correct_format_runs / total_runs | >= 0.95 |
| response_time_p50 | median seconds to complete | <= 30s |
| response_time_p95 | 95th percentile seconds | <= 60s |
| token_usage_avg | average tokens per run | tracked only |
| pass_rate | fixtures scoring above min_score | >= 0.80 |
Per-Agent Benchmark Definitions
code-reviewer
Fixtures: 6 (2 missing error handling, 2 code smell, 1 SQL injection, 1 clean code) Pass threshold: 70/100 Critical findings: error handling, injection vulnerabilities, magic numbers Non-critical findings: naming conventions, comment quality
security-reviewer
Fixtures: 8 (hardcoded secrets, injection flaws, auth bypass, safe code) Pass threshold: 75/100 Zero tolerance: must find all HIGH/CRITICAL security issues Acceptable miss: LOW severity cosmetic issues only
verifier
Fixtures: 4 (passing build, type errors, failing tests, lint errors) Pass threshold: 80/100 Critical: must correctly identify PASS vs FAIL state Scoring focus: verdict accuracy over prose quality
sleuth (bug investigator)
Fixtures: 5 (null pointer, race condition, wrong logic, correct code) Pass threshold: 65/100 Critical: must identify root cause, not just symptom Scoring focus: root cause analysis depth
Baseline Management
Baseline File Format
json{ "agent": "code-reviewer", "created_at": "2026-03-26T00:00:00Z", "commit": "abc1234", "scores": { "missing-error-handling": 88, "sql-injection": 92, "clean-code": 95, "code-smell-nesting": 79, "magic-numbers": 82, "dead-code": 76 }, "aggregate": { "average": 85.3, "min": 76, "max": 95, "pass_rate": 1.0 } }
Baseline Lifecycle
Create baseline → Make changes → Run benchmark → Compare → PASS (no regression) → Update baseline → FAIL (regression) → Fix and rerun
CI Integration
GitHub Actions Example
yamlname: Agent Benchmark on: push: paths: - '.claude/agents/**' - '.claude/skills/**' jobs: benchmark: runs-on: ubuntu-latest steps: - uses: actions/checkout@v4 - name: Run agent benchmarks run: node ~/.claude/benchmarks/run.mjs --compare - name: Comment PR with results if: github.event_name == 'pull_request' uses: actions/github-script@v7 with: script: | const results = require('./benchmark-output.json') github.rest.issues.createComment({ issue_number: context.issue.number, body: formatBenchmarkResults(results) }) - name: Fail on regression run: | node ~/.claude/benchmarks/run.mjs --check-regression # Exits non-zero if regression > 10 points on any fixture
Benchmark Authoring Guide
Writing a Good Fixture
A good benchmark fixture is:
- Realistic - Code that could exist in a real project
- Focused - Tests one specific thing the agent should find
- Unambiguous - The ground truth is objectively correct
- Minimal - No unnecessary noise that could confuse the agent
Example: Good Fixture (code-reviewer)
typescript// fixtures/code-reviewer/missing-error-handling.ts // BENCHMARK: Agent must find missing error handling in fetchUser async function fetchUser(id: string) { const response = await fetch(`/api/users/${id}`) const data = await response.json() return data } export default fetchUser
Ground truth:
json{ "required_findings": [{ "severity": "HIGH", "description_contains": ["error handling", "network", "try"], "location_hint": "fetchUser" }], "required_verdict": "FAIL", "min_score": 70 }
Example: Bad Fixture (too complex)
Do not create fixtures with 10 different issues. The agent may find 7, miss 3, and you cannot tell if the misses are regressions or noise. One fixture = one primary concern.
Integration with Canavar
When a benchmark run produces a regression, log it to the Canavar error ledger:
bashnode ~/.claude/hooks/dist/canavar-cli.mjs errors
Canavar cross-training means a regression in code-reviewer will inject a warning into all producer agents that use code-reviewer output, preventing cascading quality failures.
Quick Reference
bash# Before changing an agent: node ~/.claude/benchmarks/run.mjs --agent code-reviewer --save-as before # After changing the agent: node ~/.claude/benchmarks/run.mjs --agent code-reviewer --compare before # Full regression check: node ~/.claude/benchmarks/run.mjs --compare --fail-on-regression # Update baselines after confirmed improvement: node ~/.claude/benchmarks/run.mjs --baseline update
Remember: A benchmark suite that is never run is decoration. Run benchmarks before every agent change. Protect quality proactively, not reactively.
