# Evoclawbench

> EvoClawBench - Benchmark for evaluating LLM agent skill evolution capabilities. Tests whether agents can identify repeating patterns, create reusable skills, and effectively reuse them across sub-problems.

- Skill: `pzy2000/evoclawbench` (Agent Skill, multi-file: 673 files)
- Install (CLI): `npx skillmds@latest add pzy2000/evoclawbench`
- Raw SKILL.md: https://api.skillmd.com/api/skills/pzy2000/evoclawbench/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: pzy2000 (https://skillmd.com/u/pzy2000)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/pzy2000/evoclawbench

---


# EvoClawBench

An open-source benchmark for evaluating LLM agents' ability to create and reuse skills (auto-evolution) at runtime.

## Quick Start

```bash
uv run scripts/benchmark.py --model anthropic/claude-sonnet-4 --runtime nanobot --mode both
```

## Key Concepts

- **Baseline mode**: Agent is forbidden from creating skills
- **Evolution mode**: Agent is encouraged to create reusable skills
- **fail2pass ratio**: Measures the benefit of skill creation