Grammar Search for Multi-Agent Systems

TL;DR

This paper introduces a structured, component-based search framework for multi-agent systems that outperforms LLM-based methods on several benchmarks, offering cost efficiency and interpretability.

Contribution

It presents a novel fixed-component search approach that surpasses prior LLM-based methods in efficiency and system modularity.

Findings

Outperforms prior methods on 4 out of 5 benchmarks

More cost-efficient search process

Generates modular, interpretable multi-agent systems

Abstract

Automatic search for Multi-Agent Systems has recently emerged as a key focus in agentic AI research. Several prior approaches have relied on LLM-based free-form search over the code space. In this work, we propose a more structured framework that explores the same space through a fixed set of simple, composable components. We show that, despite lacking the generative flexibility of LLMs during the candidate generation stage, our method outperforms prior approaches on four out of five benchmarks across two domains: mathematics and question answering. Furthermore, our method offers additional advantages, including a more cost-efficient search process and the generation of modular, interpretable multi-agent systems with simpler logic.