ADAS: When Agents Are Created By Agents

What can be more life inspiring than watching your Agents making their own Agents? Actually, many, but still LLM agents creating LLM agents is a joy of innovation.

This paper proposes a new research area called Automated Design of Agentic Systems (ADAS), which aims to automatically design powerful agentic systems by learning new building blocks and combining them in innovative ways.

Shortly:

• ADAS is a new research area that aims to automate the design of agentic systems, which are AI systems that use Foundation Models as modules to solve complex tasks.
• The authors propose Meta Agent Search, an algorithm that uses a meta agent (a Foundation Model) to iteratively program new agents in code, based on an ever-growing archive of previous discoveries.
• Experiments across multiple domains, including coding, science, and math, show that Meta Agent Search consistently discovers agents that outperform state-of-the-art hand-designed agents.
• The discovered agents exhibit strong transferability, demonstrating their robustness and generality.
• The authors argue that defining agents in code offers better interpretability, making debugging easier and enhancing AI safety.
• ADAS has the potential to accelerate the development of powerful AI systems and contribute to the creation of Artificial General Intelligence (AGI).
• The paper highlights the importance of addressing safety concerns and developing safe ADAS algorithms to ensure that the technology is used responsibly.

The Paper: https://arxiv.org/pdf/2408.08435

Source code: https://github.com/ShengranHu/ADAS

At PREDICTif we vigilantly monitor emerging GenAI technologies on your behalf. Our team of young, talented data science engineers stands ready and happy to help you unleash the power of this new technological wave.