Dynamically Generate new Agent Systems

[yeandy]

Hi folks, I've been working on a project to dynamically generate new AutoGen agent systems. Based on the ADAS paper here, the idea is to use a meta-agent to generate novel agent systems. We give the meta-agent code examples of some well-known agent systems, such as Chain-of-thought, Chain-of-thought with self consistency (Mixture of Agents), LLM debate, or Reflection (Actor-Critic), and then guide it to write new AutoGen-Core code for a novel agent system.

The motivation for this project is that there is a creativity and bandwidth limit on what we can craft manually; having an agent to do this may yield surprising architectures that could work better than the current known ones. Ideally, the meta-agent will be able to design, using AutoGen-Core building blocks, a new (multi-)agent system, which if proven useful (after going through a period of testing/adoption by other developers), be incorporated into the official AutoGen-AgentChat API.

Stay tuned for a PR and a user guide those who wish to try this out. Thanks!