Controlled study finds multi-agent setups give 0.0% average gain over a single agent

bravo_abad · x · 2026-07-28

A controlled study of multi-agent coordination finds no overall gain from adding more agents when prompts, tools, and compute budgets are held fixed.

The authors test 260 configurations across five architectures: single agent, independent, centralized, decentralized, and hybrid.
They evaluate three model families on six agentic benchmarks, ranging from web browsing to software engineering and command-line tasks.
The aggregate effect of coordination is essentially zero: mean improvement over a single agent is 0.0%, with a 95% confidence interval from -58.7% to +77.2%.
The benefit is highly task-dependent. Financial analysis improves by +80.8% with a centralized orchestrator, while other tasks can get worse.
The paper’s message: multi-agent systems help only when a single agent is already struggling, and even then the gains are inconsistent.

More from coding & agent