HyperAgents: Self-Improving AI Agents from Meta

Discover HyperAgents, Facebook Research's groundbreaking open-source project featuring self-referential, self-improving AI agents that optimize for any computable task. This repository includes complete implementations of meta-agents and task-agents, supports multiple LLM providers (OpenAI, Anthropic, Gemini), and comes with Docker support, experiment logs, and domain-specific code. Learn how to set up and run these cutting-edge agents that can iteratively improve their own code. Safety warning included for executing model-generated code.

HyperAgents: Meta's Breakthrough in Self-Improving AI Agents

Facebook Research has unveiled HyperAgents, a revolutionary open-source framework for creating self-referential, self-improving AI agents capable of optimizing any computable task. This isn't just theory - it's a fully functional GitHub repository ready for you to deploy.

What Makes HyperAgents Unique?

Unlike traditional agents, HyperAgents feature a hierarchical architecture:

  • Task Agents: Handle specific domains and generate solutions
  • Meta Agents: Analyze task agent performance, identify weaknesses, and generate improved versions
  • Self-Referential Loop: Agents improve their own code through iterative evolution

Quick Setup Guide

# 1. Set API keys in .env
OPENAI_API_KEY=your_key
ANTHROPIC_API_KEY=your_key
GEMINI_API_KEY=your_key

# 2. Install dependencies
sudo dnf install -y python3.12-devel graphviz...
python3.12 -m venv venv_nat
pip install -r requirements.txt

# 3. Build Docker (optional)
docker build -t hyperagents .

# 4. Setup initial agents
bash setup_initial.sh

# 5. Run experiments
python generate_loop.py --domains <domain>

Key Features

Multi-LLM Support: OpenAI, Anthropic, Gemini ✅ Docker Containerization: Easy deployment ✅ Multiple Domains: Pre-built task environments ✅ Analysis Tools: Performance visualization scripts ✅ Experiment Logs: 2.1K+ stars, complete reproduction

File Structure Highlights

  • meta_agent.py / task_agent.py: Core agent implementations
  • generate_loop.py: Main training loop entrypoint
  • domains/: Task-specific environments
  • analysis/: Performance evaluation tools

Safety First

⚠️ Warning: HyperAgents execute model-generated code. While malicious behavior is unlikely with current models, destructive actions remain possible due to alignment limitations.

Research Paper

arXiv:2603.19461 - HyperAgents: Self-referential self-improving agents

Get Started Today

⭐ Star on GitHub and join 2.1K+ developers exploring the future of autonomous AI improvement. Perfect for AI researchers, agent developers, and anyone pushing the boundaries of what LLMs can achieve.

This project represents a major leap toward truly autonomous AI systems that don't just solve tasks, but evolve themselves to solve them better.