Putting three agents in a room does not create a team. It creates three failure surfaces and a coordination problem with better branding.
CoCoBench, a new preprint, evaluates 11 multimodal language models across 897 executable household-task instances. The benchmark isolates four coordination capabilities: allocating tasks, preserving sequential order, preventing conflicting use of a shared resource, and handing work from one agent to another.
That separation is the point. The researchers find that strong aggregate results do not mean a model is evenly competent. A system can allocate work effectively and still botch handoffs. Another can respect order until more agents or different observations change the operating environment. One victory percentage compresses those differences into a number that looks more useful than it is.
The tasks are oracle-validated and executable, which gives the benchmark more structure than a pile of model-written plans graded by another model. It also explores different coordination modes, observation inputs, and team sizes. CoCoBench remains a preprint, so the claims should be treated as early research rather than settled doctrine.
The business implication is immediate. Any company selling a swarm, agent workforce, or autonomous operations layer needs construct-level evidence. Who assigns the work? Who owns state? What happens when two agents need the same tool? How does one know a handoff succeeded? If those answers are missing, the product is not coordinated intelligence. It is parallel improvisation with a dashboard.
LaunchPad positionMulti-agent performance is not one capability. Teams need to know whether a system can allocate work, preserve order, share scarce resources, and execute handoffs without confusing aggregate success for balanced competence.
This report draws on the linked primary sources and reputable reporting. Company statements are treated as claims until independently demonstrated.