Automated Design of Agentic System

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

By: Shengran Hu, Cong Lu, Jeff Clune

Abstract

Researchers are investing substantial effort in developing powerful general-purpose agents, wherein Foundation Models are used as modules within agentic systems (e.g. Chain-of-Thought, Self-Reflection, Toolformer). However, the history of machine learning teaches us that learned solutions eventually replace hand-designed solutions. We formulate a new research area, Automated Design of Agentic Systems (ADAS), which aims to automatically create powerful agentic system designs, including inventing novel building blocks and/or combining them in new ways. We further demonstrate that there is an unexplored yet promising approach within ADAS where agents can be defined in code and new agents can be automatically discovered by a meta agent programming even better ones in code. Given that programming languages are Turing Complete, this approach theoretically enables learning any possible agentic system: including novel prompts, tool use, control flows, and combinations thereof. We present a simple yet effective algorithm named Meta Agent Search to demonstrate this idea, where a meta agent iteratively programs interesting new agents based on an ever-growing archive of previous discoveries. Through extensive experiments across multiple domains including coding, science, and math, we show that our algorithm can progressively invent agents with novel designs that greatly outperform state-of-the-art hand-designed agents. Importantly, we consistently observe the surprising result that agents invented by Meta Agent Search maintain superior performance even when transferred across domains and models, demonstrating their robustness and generality. Provided we develop it safely, our work illustrates the potential of an exciting new research direction toward automatically designing ever-more powerful agentic systems to benefit humanity.

Summary Notes

Figure: Overview of the proposed algorithm Meta Agent Search and examples of discovered agents. In our algorithm, we instruct the “meta” agent to iteratively program new agents, test their performance on tasks, add them to an archive of discovered agents, and use this archive to inform the meta agent in subsequent iterations. We show three example agents across our runs, with all names generated by the meta agent. The detailed code of example agents can be found in Appendix

Introduction

The development of powerful general-purpose agents using Foundation Models (FMs) like GPT and Claude is rapidly advancing. These agents are designed for complex tasks that require flexible reasoning and planning.

However, manually designing these systems is time-consuming and increasingly impractical as the complexity of tasks grows. This paper introduces a new research area called Automated Design of Agentic Systems (ADAS), which aims to automate the creation of these agentic systems. The core idea is that instead of relying on manual design, a meta-agent can automatically generate and refine agents in code, leveraging previous discoveries to create increasingly effective and innovative agents.

Key Concepts

Automated Design of Agentic Systems (ADAS)

• ADAS is conceptualized as an optimization framework where a search algorithm discovers and optimizes agentic systems within a vast search space.
• The approach emphasizes the use of code as the search space, allowing the discovery of novel prompts, tool uses, control flows, and their combinations.

Meta Agent Search Algorithm

• This algorithm allows a meta-agent to iteratively create new agents by coding them, evaluate their performance on tasks, and refine them based on an archive of previous discoveries.
• The meta-agent focuses on creating agents that are novel and effective, progressively improving their design through self-reflection and iterative refinement.

Experiments and Results

ARC Challenge

• Meta Agent Search was tested on the ARC logic puzzle task, a benchmark for evaluating general intelligence in AI systems.
• The algorithm discovered agents that significantly outperformed state-of-the-art hand-designed agents, demonstrating the potential of ADAS to innovate and refine agentic designs continuously.

Reasoning and Problem-Solving Domains

• The algorithm was further tested on benchmarks in reading comprehension, math, science, and multi-task problem-solving.
• The discovered agents outperformed existing hand-designed agents across these domains, particularly in reading comprehension and math, where the improvements were most pronounced.

Transferability Across Models and Domains

• The discovered agents were tested across different foundation models and domains, showing strong generalization and transferability.
• These agents maintained high performance even when transferred to different models, highlighting their robustness and adaptability.

Examples of Discovered Agents

• The paper presents several examples of agents discovered by Meta Agent Search, such as the "Divide and Conquer Agent," "Multi-step Peer Review Agent," and "Verified Multimodal Agent," each designed to tackle specific challenges in different tasks.

Discussion and Future Work

The paper emphasizes the potential of ADAS to automate the design of agentic systems, saving human effort and improving upon manual designs. Future research directions include exploring higher-order ADAS, integrating existing building blocks, and developing multi-objective ADAS algorithms. The authors also highlight the importance of safety considerations when executing model-generated code, advocating for sandbox environments to mitigate risks. Additionally, they discuss the broader implications of advancing AI capabilities, particularly through ADAS, and call for careful and responsible research practices.

Conclusion

This research introduces ADAS as a promising new direction in AI, capable of automating the invention of powerful agentic systems. The Meta Agent Search algorithm exemplifies the potential of this approach, consistently outperforming hand-designed agents and demonstrating the ability to generalize across different models and domains. The work opens up numerous avenues for future exploration, particularly in the realm of automated agent design, which could revolutionize how AI systems are developed and deployed.