A-Evolve: evolution of AI agents in 3 lines of code

A-Evolve: Framework for Automatic Evolution of LLM-Based AI Agents

The A-Evolve framework from researchers at Amazon and the University of Pennsylvania transforms a basic AI agent into a high-performance one through iterative mutations. The system modifies prompts, skills, memory, and tools, tests changes on benchmarks, and retains only improvements. Launching it requires just three lines of Python code; the process is fully autonomous and uses Git for versioning mutations.

This approach solves the problem of agent degradation in changing environments: instead of fine-tuning LLMs or manually tweaking prompts, it deploys an "evolver agent." It analyzes failures, makes edits to the agent's working directory, and reruns evaluations. This aligns with the "agentic evolution" concept from the paper Position: Agentic Evolution is the Path to Evolving LLMs (January 2026).

Architecture and How It Works

A-Evolve follows the BYOA (Bring Your Own Agent), BYOE (Bring Your Own Environment), and BYO-Algo (Bring Your Own Evolution Algorithm) models. The agent's working directory is a standard folder with files:

• manifest.yaml — configuration;
• prompts in separate files;
• skills as scripts;
• memory in vector database or log format;
• tools as callable modules.

The mutation engine edits these files: rephrases prompts, adds/removes skills, optimizes memory. After mutation, it runs benchmarks. Improvements are committed to Git, ensuring reproducibility.

Supported LLMs: Anthropic (Claude), OpenAI, Amazon Bedrock. Benchmarks: MCP-Atlas, SWE-bench Verified, Terminal-Bench 2.0, SkillsBench.

Example integration (three lines of code):

from a_evolve import Evolver
evolver = Evolver(agent_dir="./my_agent", benchmark="mcp-atlas")
evolver.evolve(generations=50)

Results on Key Benchmarks

Testing was conducted on an agent with Claude Opus 4.6. Evolution delivered significant performance gains:

| Benchmark | Baseline (%) | After Evolution (%) | Gain (p.p.) |

|-----------------------|--------------|---------------------|-------------|

| MCP-Atlas | 76.0 | 79.4 | +3.4 |

| SWE-bench Verified | 74.2 | 76.8 | +2.6 |

| Terminal-Bench 2.0 | 63.5 | 76.5 | +13.0 |

| SkillsBench | 19.7 | 34.9 | +15.2 |

On MCP-Atlas, the 79.4% score propelled the agent to first place in the overall ranking. SWE-bench Verified landed around fifth. The gains on Terminal-Bench and SkillsBench stand out, highlighting effectiveness on terminal interface tasks and skills assessments.

Benefits for Developers

• Scalability: No need for manual prompt iteration on every agent.
• Reproducibility: Git commits capture every successful mutation.
• Flexibility: Import any agent or benchmark without code rewrites.
• Efficiency: Autonomous analyze-mutate-test cycle minimizes human involvement.
• Open-source: Framework ready for production pipeline integration.

For mid/senior developers, it's a tool for automating agent A/B testing. The evolver integrates into CI/CD, triggering mutations on nightly builds.

Key Takeaways

• A-Evolve boosts a baseline Claude Opus 4.6 agent to 79.4% on MCP-Atlas — top of the ranking.
• Gains up to 15 p.p. on SkillsBench and Terminal-Bench without model fine-tuning.
• Three lines of code to start: Evolver(agent_dir, benchmark).evolve().
• Git versioning tracks all mutations comprehensively.
• Supports Anthropic, OpenAI, Bedrock — ready for multi-LLM setups.

— Editorial Team