A-EVO-Lab/a-evolve

What it solves

A-Evolve provides a universal infrastructure for the autonomous self-improvement of AI agents. It eliminates the need for manual harness engineering by allowing agents to evolve their own prompts, skills, and memory across any domain or benchmark using various evolution algorithms.

How it works

The system operates on a "file system contract" where all evolvable agent state (prompts, skills, and memory) is stored in a standard directory structure. This allows an evolution engine to mutate the agent's workspace files via LLM-driven operations without needing to know the agent's internal architecture.

The evolution process follows a five-phase loop:

1. Solve: The agent executes a batch of tasks.
2. Observe: Trajectories and benchmark feedback are collected into logs.
3. Evolve: The engine analyzes logs and mutates workspace files.
4. Gate: Mutations are validated on holdout tasks and rolled back via git if they cause regression.
5. Reload: The agent reloads the updated workspace.

Who it’s for

It is designed for AI researchers and developers working on LLM self-optimization, reinforcement learning, and agent architectures who want to test and evolve agents across diverse environments like SWE-bench or MCP-Atlas.

Highlights

• Universal Pluggability: Supports "Bring Your Own Agent" (BYOA), "Bring Your Own Benchmark" (BYOE), and "Bring Your Own Algorithm" (BYO-Algo).
• Zero Manual Engineering: Enables SOTA performance on multiple benchmarks (e.g., MCP-Atlas, SWE-bench Verified) with minimal code.
• Git-Based Versioning: Every accepted mutation is git-tagged for full reproducibility and audit trails.
• Reference Algorithms: Includes four built-in evolution strategies: adaptive_evolve, adaptive_skill, skillforge, and guided_synth.