Abstract
Large language models, employed as multiple agents that interact andcollaborate with each other, have excelled at solving complex tasks. The agentsare programmed with prompts that declare their functionality, along with thetopologies that orchestrate interactions across agents. Designing prompts andtopologies for multi-agent systems (MAS) is inherently complex. To automate theentire design process, we first conduct an in-depth analysis of the designspace aiming to understand the factors behind building effective MAS. We revealthat prompts together with topologies play critical roles in enabling moreeffective MAS design. Based on the insights, we propose Multi-Agent SystemSearch (MASS), a MAS optimization framework that efficiently exploits thecomplex MAS design space by interleaving its optimization stages, from local toglobal, from prompts to topologies, over three stages: 1) block-level (local)prompt optimization; 2) workflow topology optimization; 3) workflow-level(global) prompt optimization, where each stage is conditioned on theiteratively optimized prompts/topologies from former stages. We show thatMASS-optimized multi-agent systems outperform a spectrum of existingalternatives by a substantial margin. Based on the MASS-found systems, wefinally propose design principles behind building effective multi-agentsystems.