AI Agents That Talk to Each Other Can Actually Get Worse

We're increasingly surrounded by AI systems that talk to each other — recommendation engines, navigation apps sharing traffic, chatbots coordinating on a task. The common assumption is that more communication means better decisions, the way a group of people usually beats one person guessing. A new paper from researcher Xuening Wu argues that's not guaranteed. Communication can push AI agents into agreement while making their collective answer worse.

The paper's first finding is about how easily we can be fooled by simple scorecards. Two setups can look identical on paper — same statistics, same numbers — yet one produces a good result and the other a bad one. In the study's test, just changing which direction messages traveled took accuracy from 72.6% up to 91.2%, or down to 65.9%. The standard shortcut for predicting how well a network will do misses this direction entirely. The author's fix: use a small set of labeled examples (data with known right answers) kept apart from the real test, which predicted multi-round results to within about half a percentage point on two made-up tasks, then held up on image tasks including handwritten digits.

The second finding hits closer to home. When many connected systems share the same built-in bias, the group's average accuracy can improve while specific communities are harmed — or while the overall group vote gets worse. Think of a fleet of delivery robots all trained on the same skewed data: on average they might look sharper while consistently failing one neighborhood. Adding calibration limits reduced the measured harm while keeping much of the benefit, but it did not guarantee protection.

The honest caveat, in the author's own words: broad transfer to other settings and practical superiority are left "open." This is an early mathematical warning, not a product. Its value is a rule of thumb worth remembering as AI agents start coordinating on your calendar, your shopping, and your care.