Summary

A-Evolve offers a universal, open-source infrastructure for automating the evolution of AI agents using any evolutionary algorithm. It targets researchers and developers aiming to achieve state-of-the-art (SOTA) performance with minimal manual harness engineering, significantly accelerating agent improvement across diverse domains.

How It Works

The core mechanism centers on the "Agent Workspace," a standardized file system contract (manifest, prompts, skills, memory) enabling the evolution engine to mutate agent components externally. A five-phase loop (Solve, Observe, Evolve, Gate, Reload) drives the process, with mutations applied to workspace files, validated, and git-tagged for reproducibility.

Quick Start & Requirements

Installation is via pip install -e ".[all,dev]". Key dependencies include Python and PyTorch (implied). The project includes built-in seed workspaces and benchmark adapters, with guides available for specific demos. The official paper is on arXiv: 2602.00359.

Highlighted Details

• Achieved SOTA on four benchmarks (MCP-Atlas #1, SWE-bench Verified ~#5, Terminal-Bench 2.0 ~#7, SkillsBench #2) with zero manual harness engineering.
• Enables agent evolution with "3 lines of code" for basic setups.
• Defines a clear "Agent Workspace" file system contract for state management.
• All evolutionary mutations are git-tagged for complete reproducibility.
• Highly pluggable architecture supports custom agents, benchmarks, algorithms, and LLMs.

Maintenance & Community

A-Evolve is an open-source project welcoming community contributions. Users can report issues, submit PRs, and join the Discord server for collaboration. Starring the repository supports the research.

Licensing & Compatibility

Released under the permissive MIT License, allowing broad use, modification, and distribution, including for commercial purposes.

Limitations & Caveats

The README does not detail specific hardware requirements or known bugs. Achieving optimal results may require significant computational resources for extensive evolution cycles. LLM-driven mutations can inherently introduce unpredictability.