agent-config-arena

> Everyone shares their CLAUDE.md. Nobody benchmarks them. Until now.

[](LICENSE) [](https://nodejs.org) [](CONTRIBUTING.md)

| Config | Pass Rate | Avg Tokens | Avg Time | Avg Cost | Score | |--------|-----------|------------|----------|----------|-------| | token-efficient | 88% | 208k | 73.7s | $0.28 | 44 | | workflow-heavy | 86% | 205k | 90.0s | $0.31 | 40 | | baseline (no config) | 88% | 201k | 112.9s | $0.33 | 36 |

> Tested on 8 real coding tasks (REST API, refactoring, bug fix, CLI tool, data pipeline, test coverage, TS migration, performance optimization). Full results in LEADERBOARD.md.

• "Token-efficient" config is fastest and cheapest, but doesn't actually use fewer tokens. It wins on speed (35% faster than baseline) and cost, not token count.
• Zero config (baseline) uses the fewest tokens. Adding instructions makes the model *more* verbose, not less.
• "Workflow-heavy" (plan-first, TDD) has the lowest pass rate. More structure doesn't mean more correct.
• All 3 configs failed the data pipeline task. The hardest tasks expose config limitations equally.
• The biggest gap is speed, not accuracy. Pass rates are within 2%, but time ranges from 73s to 113s.

---

A neutral, open-source benchmark that tests different coding agent configurations on the same real coding tasks -- then publishes the results.

Not a model benchmark. SWE-bench tests models. We test *configs*. Not a tool collection. awesome-claude-code collects tools. We *evaluate* them. Not a single config. claude-token-efficient ships one config. We pit configs *ag