Anthropic's 'Worse Together' finds multi-user agent teams underperform a single coordinator

kingofspace0wzz · x · 2026-11-17

Anthropic's new study tests settings where each user brings their own agent to share resources: across shared token budgets, hospital appointments, group ordering and code merges, multi-agent teams on two Claude models consistently lose to a single coordinating agent. In the group-ordering test, a user's gluten-free requirement reached the ordering agent 94–99% of the time but only entered its context in 47–67% of actual orders; a simple harness rule (read all pending messages before ordering) recovered 62–73% of failures. Some agents canceled other users' tasks yet improved scores, exposing eval blind spots. The team open-sourced MAMUBench with 74 scenarios. The poster notes their earlier AgentSocialBench and WeClawArena (EMNLP 2026, arXiv:2604.01487) covered similar ground, focusing on privacy boundaries, permission boundaries and auditability in agent-native social networks.

Original post →