howler-agents

Group-Evolving AI Agents (GEA) — groups of AI agents evolve together by sharing experience across lineages.

Implements the GEA system from arXiv:2602.04837, achieving 73.3% on SWE-bench Lite.

Quick Start

Initialize any repository with howler-agents skills, MCP server, and agent definitions:

npx howler-agents init

This sets up:

• 9 Claude Code skills (/howler-agents, /howler-agents-wiggam, /howler-init, etc.)
• 4 agent definitions (coordinator, evaluator, reproducer, actor)
• MCP server registration in .mcp.json
• Local .howler-agents/ directory

Then use the slash commands in Claude Code:

/howler-init                           # Seed hive-mind with repo knowledge
/howler-agents Fix the auth bug        # Solve a task with collective intelligence
/howler-agents-wiggam Fix all tests \
  --completion-promise "ALL TESTS PASSING"  # Iterate until done

Init Options

| Flag | Default | Description | |------|---------|-------------| | --command | npx | MCP server command: npx, uvx, or howler-agents | | --overwrite | off | Replace existing skill/agent files | | --skip-skills | off | Skip installing Claude Code skills | | --skip-agents | off | Skip installing agent definitions | | --skip-mcp | off | Skip MCP server registration |

Other Commands

npx howler-agents serve          # Start MCP server (stdio transport)
npx howler-agents evolve         # Run evolution locally
npx howler-agents install        # Register with AI coding tools

How It Works

This is a thin Node.js wrapper that auto-installs the Python howler-agents-core package (via uv or pip) on first use, then proxies all commands to the Python CLI.

Documentation

• Getting Started
• Skills Reference
• Architecture
• Paper Results

License

MIT