Automated Design of Agentic Systems (ADAS)

When agents are designing agents

Generative AI (GenAI) is evolving rapidly and the concept of agentic systems has gained significant attention. These systems are essentially autonomous agents designed to perform specific tasks, often by interacting with complex environments. However, designing these agents typically requires significant manual effort and expertise.

The recent paper "Automated Design of Agentic Systems (ADAS)" introduces an intruiging approach to automate the design of such systems, offering a glimpse into the future of fully automated agentic AI.

1. Introduction to ADAS

The core premise of ADAS is to automate the creation of agentic systems through a process called Meta Agent Search. The idea is to use a meta-agent—an agent designed to create other agents by writing and refining code iteratively. This meta-agent operates in the code space, generating new agents by combining existing components or creating entirely new ones from scratch. The goal is to discover novel and powerful agentic systems that can outperform manually designed agents.

2. Meta Agent Search: The Heart of ADAS

Meta Agent Search is the key mechanism that enables ADAS. It functions as follows:

1. Initial Design: The meta-agent starts by designing a simple agent using basic building blocks. These blocks can be anything from existing functions and libraries to custom code snippets created by the meta-agent itself.

2. Evaluation: The newly created agent is then evaluated on a set of predefined tasks. The performance of the agent is measured, and detailed logs are generated, capturing the agent's behavior and decision-making process.

3. Iteration and Refinement: Based on the evaluation, the meta-agent refines the design by tweaking the code, adding new components, or even discarding ineffective parts. This iterative process continues, with each iteration producing more sophisticated agents.

4. Exploration and Exploitation: The search process balances exploration (trying out new designs) and exploitation (refining known good designs) to ensure that the meta-agent discovers a wide variety of potential solutions while also honing in on the most promising ones.

3. Key Features and Innovations