Multi Agent Coordination Bench (MAC Bench) is a benchmark of twelve long-horizon tasks in simulated environments, designed to test whether teams can coordinate - not merely whether adding agents increases output. Tasks are built so that each agent’s work depends on teammates’ contributions, spanning domains such as air-traffic control, drone rescue, software development, and a Mars sample-return mission. The benchmark distinguishes shared tasks, where agents have the same information and capabilities, from distributed tasks, where the environment divides what agents can observe or do. All included tasks have a common team goal; competitive tasks are excluded because opponents would make performance and coordination gains harder to compare. Each task is tested with one agent, multiple agents without coordination, and multiple agents able to communicate, separating the benefit of adding agents from the benefit of teamwork.
Across models from OpenAI, Anthropic, and Google, teams of eight score an average of 0.42, compared with 0.35 for a single agent. Yet eight agents without coordination score 0.44, indicating that communication generally fails to improve results; only Opus 5.5 gains from messaging. The leaderboard also reveals that similar solo performance can translate into markedly different teamwork: Opus 5.5 reaches 0.53 in teams of eight, while GPT-6.1 Sol scores 0.41. Scaling is costly and inefficient: an eight-agent team costs about nine times as much as one agent and reaches its score only about twice as fast. The benchmark therefore identifies room to improve not just task performance, but also coordination and the efficiency of multi-agent compute.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.