This repository introduces Automated Design of Agentic Systems (ADAS), a new research area focused on automatically creating agentic systems. It targets researchers and practitioners seeking to automate the discovery of novel agent architectures and functionalities, offering a path to more powerful and adaptable AI agents.

How It Works

ADAS employs a "meta" agent that iteratively generates agent designs in code. This meta-agent learns from past discoveries, programming new agents that incorporate novel building blocks or combinations. The approach is demonstrated via Meta Agent Search, which explores different domains by having the meta-agent program and refine agent designs.

Quick Start & Requirements

• Install: conda create -n adas python=3.11, conda activate adas, pip install -r requirements.txt.
• Prerequisites: OpenAI API key (set via export OPENAI_API_KEY="YOUR KEY HERE").
• Running: Navigate to domain folders (e.g., _arc, _drop) and run python {DOMAIN}/search.py.
• Customization: Modify evaluate_forward_fn(), prompts, and add basic functions in search.py for new domains.
• Docs: Tweet (for context on awards).

Highlighted Details

• Outstanding Paper at ICLR 2025 and NeurIPS 2024 Open-World Agent Workshop.
• Demonstrates agents inventing novel and powerful agent designs.
• Code is self-contained within domain-specific folders.

Maintenance & Community

• Primary contributor: ShengranHu.
• No explicit community links (Discord/Slack) or roadmap provided in the README.

Licensing & Compatibility

• License not specified in the README.
• Compatibility for commercial or closed-source use is undetermined.

Limitations & Caveats

The code executes untrusted model-generated code, posing a potential risk of destructive actions due to model limitations, even if malicious intent is unlikely in current settings. Users must acknowledge and accept these risks.