Introduction

The study, titled 'Worse Together: How Performance Breaks Down in Multi-User Multi-Agent Teams,' found that teams of AI agents serving different users consistently underperform a single coordinating agent when sharing limited resources.

The Research Environment

The researchers tested five advanced models across 77 scenarios, including shared API-token budgets, clinic scheduling, personal-assistant bookings, and code-merge queues. They also introduced MAMUBench, a new benchmarking framework designed for standardized evaluation across three key environments.

The Results

The team evaluated five advanced models across 77 scenarios. Peer-to-peer teams reached only 12-30% of optimal group outcomes across the different environments. Single-agent coordinators landed between 32-64% of optimal. Silent teams fared worst, with success rates falling to as low as 2-7% in certain scenarios. Turning on communication did help, but only modestly.

The Failure Modes

The study identifies specific failure modes behind the decline. Agents stalled rather than acting, and they overrode actions taken by their peers. Both behaviors showed up even when communication channels were available. The personal-assistant environment produced one of the starkest comparisons, where single-agent coordinators fulfilled targeted user requests twice as often as multi-agent teams.

The Implications

The study's most pointed conclusion targets a popular assumption. According to the researchers, the advantages often credited to multi-agent strategies frequently come from extra computational resources, not genuine collaboration. The paper also argues that decentralization can amplify conflicts among agents with competing claims, rather than ease them.

The Fixes

The study provides practical fixes tailored to each environment. These include improving agent communication, reducing the number of agents, and increasing the computational resources available to the agents.