autoevolve

Multi-agent research competition orchestrator for autoresearch. Run parallel AI agents with different strategies and cross-pollinate winning ideas.

Install

pip install autoevolve

Usage

# Initialize a 3-agent competition
autoevolve init --agents 3 --tag mar15

# Check who's winning
autoevolve status
autoevolve leaderboard --detailed

# Spread winning ideas to all agents
autoevolve pollinate

# Export results
autoevolve export --format json -o evolve-results.json

How It Works

1. init creates a git worktree per agent in a sibling directory, each with a different research strategy
2. Each agent works independently in its worktree directory using autojudge + autosteer
3. leaderboard ranks agents by best val_bpb with keep rate tracking
4. pollinate writes the leader's best experiments to evolve-hints.md in each agent's worktree
5. Agents incorporate hints and continue competing
6. cleanup removes worktrees, branches, and config when done

Built-in Strategies

Strategies are assigned round-robin. With 3 agents, you get 3 different strategies competing.

Commands

Requirements

• Python >= 3.10
• A git repository with autoresearch set up
• Multiple compute environments (one per agent)

License

MIT