ANALYSIS OF ARCHITECTURAL COMPONENTS OF MULTI-AGENT SYSTEMS
TL;DR
It is found that on the examined tasks the multi-agent mode underperforms the single-agent baseline in quality and a pattern is discovered: the integration of three verification loops produces a self-diagnosis capability not observed in any of the examined systems individually.
Abstract
The aim of this work is to analyze the architectural components of LLM-based multi-agent systems. The research method involves a comparative analysis of MetaGPT, Generative Agents, AutoGen, OrgAgent, MIRIX, and ChatDev architectures against a uniform set of criteria (role organization, memory, coordination, structured outputs, communication, quality control), followed by integration of the components into a single system and experimental evaluation on two task types in single-agent and multi-agent modes. It is found that on the examined tasks the multi-agent mode underperforms the single-agent baseline in quality (scores of 3,0 and 4,5 versus 9,0 on a ten-point scale). The architectural cause is identified: a mismatch between task type and the preconfigured action chain. A pattern is discovered: the integration of three verification loops – a failure detector, a compliance check, and a decision-making loop – produces a self-diagnosis capability not observed in any of the examined systems individually. The three loops jointly identified result unreliability and refused to deliver it. This property differs from known multi-agent systems diagnostic approaches where analysis is performed after task completion. The results may be applied when designing multi-agent systems for tasks where the cost of error is high.