Why Letting AI Agents Chat Can Make Them Worse at Their Jobs

Companies are increasingly building AI systems that work in teams — multiple AI agents chatting, debating, and critiquing each other's work. The idea seems smart: more minds should mean better results. But a new paper shows this teamwork can backfire. When agents are shown each other's complete answers, they rapidly converge to the same idea, usually the first solution they saw. The researchers call this the 'interaction tax' — the price you pay when communication erases diversity.

To test this, they ran 11 different problem-solving tasks where AI agents had to find the best answer, using different popular AI models. When agents worked independently and only combined results at the end, the final solution was as good as or better than the chattier teams. The talkative teams simply collapsed into groupthink — every agent agreeing with the first answer, even when it was wrong. That's not just a quality problem, it's a money problem, because running multiple agents in conversation costs extra computing power without improving results.

The researchers also looked at when critique helps. They found that AI agents can only fix mistakes when the rule being broken is easy to spot — like a clear formatting error. For subtle logic mistakes, critiquing doesn't help at all. This means we can't just assume 'more interaction equals smarter.' Instead, the study says, what matters most is the right information being shared at the right time — not the number of agents.

For anyone using AI tools, the takeaway is simple and somewhat comforting: the best AI team isn't necessarily the one that talks the most. Sometimes, letting each AI think for itself and then combining ideas gets better results for less money.