Fewer than a third of multi-agent team actions help finish tasks in AgentWorld

A post about the AgentWorld paper says fewer than a third of a multi-agent team’s actions help finish its task. AgentWorld puts 3 to 20 LLM agents into a game sandbox for tasks lasting 50-plus rounds. The post says Gemini 3 Flash had the highest task success at 52.0%, while coordination tasks had 12% success. It identifies communication breakdowns, role confusion and lost shared plans as common failures.