AI agent teams face communication breakdowns in AgentWorld benchmark

AgentWorld's researchers tested teams of 3 to 20 AI agents on game-sandbox tasks spanning 50-plus interaction rounds. They report that Gemini 3 Flash had the highest task success at 52%, while communication breakdowns, role confusion and lost shared plans were recurring failures. A DAIR.AI summary says the best score for the share of actions that contributed to an outcome was 0.320, fewer than a third.