The motivating puzzle

Multi-agent systems are appealing in theory. They promise modular reasoning, distributed workload, and emergent coordination, with agents specialized in planning, coding, reviewing, tool use, and verification. Yet in practice their performance is inconsistent. Across tasks like code generation, web interaction, and software simulation, many multi-agent systems perform worse than a strong single agent, or even worse than simple best-of-N sampling (just running one model several times and picking the best answer).

The sharpest finding: these failures cannot be blamed on the model. In several cases the very same model in a single-agent setup outperforms the multi-agent version. That tells you the problem is architectural, in communication, coordination, and workflow orchestration, not a matter of the model being too weak. The paper's framing line is that coordination, not capability, is what breaks multi-agent systems.