New AI Teamwork Trick Lets 50 Robots Work Without a Boss

AI programs are increasingly being let loose to do real work — searching buildings, sorting warehouse orders, answering customer questions. But when many of these 'agents' (software that can act on its own) work together, most systems still run everything through one central coordinator, like a manager who has to read every single message. That works fine with five helpers. With fifty, the manager becomes a traffic jam, and the whole team slows down.

A research team's fix, called Proxifield, throws out the manager. Instead, each AI decides who to talk to based on 'semantic proximity' — basically, how closely someone else's current work overlaps with its own. Four simple signals drive the choice: is this agent directly addressed, does it need that information, are their plans aligned, and do they know different useful things. No retraining required, no boss assigning tasks. Think of coworkers in a big office drifting toward the people who actually have the answers they need.

The results are striking. The researchers tested their approach on drone search-and-rescue missions and on a group-reasoning test called HiddenBench. At five agents, their method beat the centralized version by 5.4%. At twenty-five agents the gap jumped to 53%, and at fifty it hit 59.5%. Equally important: when agents permanently broke down, their system kept 73.6% of its full performance, compared with 58.3% and 38.8% for the alternatives. Losing a teammate no longer cripples the team.

One honest catch: this is an academic paper, not a product you can use, and the biggest gains showed up only with enormous, expensive AI models (397 billion parameters — roughly the largest class available). The concept, though, points at where things are heading. Firefighting drones, delivery fleets, hospital scheduling bots and customer-service teams will all need a way to coordinate without a single point of failure. This paper is a blueprint for that, and it suggests the winning design looks less like a corporation and more like a flock.