evoclawbench

name	evoclawbench
description	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.

EvoClawBench

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

Quick Start

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