AI Agents Fail at Teamwork — New Test Shows Why

Imagine two robots navigating a room where one must jump in front of a laser beam so the other can roll through safely. That's the setup the researchers gave AI agents. The twist: everyone gets a reward if the team succeeds, but individuals also get points for simply escaping alone. This creates a sneaky temptation to abandon your teammate.

The researchers invented a way to mathematically certify what a task actually demands. They turn the rules into logical equations and check: can this task be completed without cooperation, or is teamwork absolutely required? Think of it like a referee who can prove whether a play “needs a pass” or could be a solo drive to the basket.

Then came the surprise. They trained five different AI learning systems on teams of tasks. When cooperation was optional, diverse training (variety of tasks) made the group smarter overall. But when cooperation was mandatory, the agents almost never finished together. They kept taking solo exits that gave them partial rewards — raking in points for leaving the team hanging.

The researchers liken it to getting paid for running out of a burning building while leaving your partner inside. The reward system rewards partial effort, but real cooperation demands finishing as a team. Their certification method reveals exactly where that gap is, so future designers can build AI incentives that actually reward working together.