Hyperagents Self Improving

---
name: hyperagents-self-improving
description: Self-referential self-improving agents from Meta Research that optimize for any computable task using meta-agents and task-agents in a recursive loop
triggers:
  - set up hyperagents
  - run self-improving agent loop
  - configure meta agent for domain
  - use hyperagents to optimize a task
  - implement self-referential agent
  - run generate loop with hyperagents
  - hyperagents experiment setup
  - facebookresearch hyperagents
---

# HyperAgents Self-Improving Agents

> Skill by [ara.so](https://ara.so) — Daily 2026 Skills collection.

HyperAgents is a Meta Research framework for self-referential, self-improving agents that recursively optimize themselves for any computable task. A **meta-agent** proposes code changes (diffs) to improve a **task-agent**, which is then evaluated on a target domain. The loop continues, progressively improving agent performance.

## Architecture Overview

generate_loop.py └── meta_agent.py ← proposes improvements (diffs) to task agent code └── task_agent.py ← executes tasks in a target domain └── agent/ ← foundation model wrappers (OpenAI, Anthropic, Gemini) └── domains/ ← domain-specific evaluation code └── run_meta_agent.py ← helper to run meta agent and get diffs

The meta-agent reads the current task-agent source, generates improved versions, applies diffs, and evaluates the new agent. This is repeated in a loop.

---

## Installation

### 1. System Dependencies (Fedora/RHEL)
```bash
sudo dnf install -y python3.12-devel
sudo dnf install -y graphviz graphviz-devel cmake ninja-build \
    bzip2-devel zlib-devel ncurses-devel libffi-devel

2. Python Environment

python3.12 -m venv venv_nat
source venv_nat/bin/activate
pip install -r requirements.txt
pip install -r requirements_dev.txt

3. API Keys

Create a .env file in the project root:

OPENAI_API_KEY=your_openai_key_here
ANTHROPIC_API_KEY=your_anthropic_key_here
GEMINI_API_KEY=your_gemini_key_here

4. Initial Agent Setup

bash ./setup_initial.sh

5. (Optional) Docker

docker build --network=host -t hyperagents .

⚠️ Safety Warning: HyperAgents executes untrusted, model-generated code. Run in an isolated environment (Docker, VM, or sandboxed container). Never run on a production system.

Key Commands

Run the Main Loop

# Basic run on a domain
python generate_loop.py --domains <domain>

# Examples:
python generate_loop.py --domains coding
python generate_loop.py --domains math
python generate_loop.py --domains reasoning

Run the Meta Agent Standalone

python run_meta_agent.py

Extract Experiment Logs

# Combine multi-part ZIP archive
zip -s 0 outputs_os_parts.zip --out unsplit_logs.zip
unzip unsplit_outputs.zip

Outputs are saved in outputs/ by default.

Core Files and Usage