HyperFlow (Python)

⚠️ EXPERIMENTAL: This project is currently in an experimental phase and is not recommended for production use.

Self-improving agent framework powered by LangChain and LangGraph.

Inspired by HyperAgents (Meta Research, 2026).

What it does

HyperFlow runs an evolutionary self-improvement loop where a MetaAgent rewrites its own source code (including prompts, tools, and logic) to make it better at solving tasks. It is self-referential: the mechanism that improves the agent is itself part of the editable code. Each generation:

1. Select a parent generation from the archive
2. MetaAgent reads past evaluation scores and edits the source code
3. Evaluation scripts run in a sandbox to score the new agent
4. Better agents are added back to the archive for future generations

Important

This framework is currently in an Experimental state. See Limitations for more information.

The TaskAgent gets better over generations without manual intervention.

New here? Read docs/concepts.md for a detailed explanation of every concept with examples.

Installation

# Install from PyPI
pip install hyperflow-ai

# Or install from source for development
pip install -e .

Requirements

• Python 3.11+
• At least one LLM provider API key (e.g. OPENAI_API_KEY, ANTHROPIC_API_KEY)

Quick Start

# Set your API key
export OPENAI_API_KEY="sk-..."

# Run the bash example (single eval)
cd examples/bash
python run.py

# Run with evolutionary loop
python run.py evolve

Project Structure

hyperflow/
  __init__.py           # Public API re-exports
  agent/
    base_agent.py       # Abstract AgentSystem base class
    llm.py              # Multi-provider LLM factory
    llm_with_tools.py   # LangGraph ReAct chat loop
    meta_agent.py       # MetaAgent (mutation operator)
    task_agent.py       # TaskAgent (task solver)
    tool_registry.py    # Tool registration
  core/
    ensemble.py         # Best-of-archive ensemble
    generate_loop.py    # Main evolutionary loop
    select_parent.py    # Parent selection strategies
  domains/
    base.py             # Domain/DomainTask/EvalResult interfaces
    evaluators.py       # Static, LLM judge, human evaluators
    harness.py          # Evaluation harness
    report.py           # Report generation
  prompts/
    llm_judge.py        # LLM judge prompt template
    meta_agent.py       # MetaAgent prompt template
    task_agent.py       # TaskAgent prompt template
  tools/
    __init__.py         # get_framework_tools()
    bash.py             # Bash shell tool
    editor.py           # File editor tool
  utils/
    archive.py          # JSONL archive CRUD
    common.py           # JSON extraction, file helpers
    constants.py        # Shared constants
    docker.py           # Docker container management
    executor.py         # Local/Docker execution
    git.py              # Git operations
examples/
  bash/                 # Bash command generation
  calculator/           # Buggy tool fix demo
  factcheck/            # True/false classification
  git_evolution/        # Git-based evolution with patches
  paper_review/         # Paper accept/reject prediction
  scoring/              # Math grading self-improvement

Supported Models

from hyperflow import MODELS

# Available model presets
MODELS["OPENAI_GPT4O"]       # "openai/gpt-4o"
MODELS["OPENAI_GPT4O_MINI"]  # "openai/gpt-4o-mini"
MODELS["OPENAI_O3"]          # "openai/o3"
MODELS["OPENAI_O4_MINI"]     # "openai/o4-mini"
MODELS["CLAUDE_SONNET"]      # "anthropic/claude-sonnet-4-5-20250929"
MODELS["GEMINI_PRO"]         # "gemini/gemini-2.5-pro"
MODELS["OLLAMA_LLAMA3"]      # "ollama/llama3"

Or use any "provider/model-name" string.

Environment Variables

Variable	Description
OPENAI_API_KEY	OpenAI API key
ANTHROPIC_API_KEY	Anthropic API key
GOOGLE_API_KEY	Google Gemini API key
OLLAMA_BASE_URL	Ollama server URL (default: http://localhost:11434)
HYPERFLOW_MODEL	Default model for examples (e.g. openai/gpt-4o)

Examples

Single Evaluation

cd examples/bash && python run.py
cd examples/factcheck && python run.py
cd examples/paper_review && python run.py

Evolutionary Self-Improvement

cd examples/bash && python run.py evolve
cd examples/factcheck && python run.py evolve
cd examples/scoring && python run.py
cd examples/calculator && python run.py
cd examples/git_evolution && python run.py

Git-Based Evolution (with patches)

cd examples/git_evolution && python run.py        # 2 generations
cd examples/git_evolution && python run.py 5      # 5 generations
cd examples/git_evolution && python run.py --reset # start over

License

MIT

Citation

If you use this framework in your research, please cite the original HyperAgents paper:

@misc{zhang2026hyperagents,
      title={Hyperagents}, 
      author={Jenny Zhang and Bingchen Zhao and Wannan Yang and Jakob Foerster and Jeff Clune and Minqi Jiang and Sam Devlin and Tatiana Shavrina},
      year={2026},
      eprint={2603.19461},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2603.19461}, 
}