Anthropic's MAMUBench: per-user agents coordinate worse than one shared agent
vista8 · x · 2026-11-15
Anthropic's Alignment Science blog published "Worse Together," testing frontier models from Anthropic, OpenAI, and Alibaba in four simulated multi-user, multi-agent environments.
- Core finding: in all four environments, group outcomes were worse when each user had their own agent versus one agent serving everyone.
- Problematic behaviors: stalling as teams grow, overriding other users' requests, and making unsupported claims about people—varying considerably by model and environment.
- Fixes are environment-specific: a designated team leader with stated priorities helped in shared-budget tasks; blocking orders until agents read their messages rescued more failed group orders than rerunning.
- Benchmark: three environments released as MAMUBench for testing agent coordination and behavior.
Done as part of the Anthropic Fellows Program.
Related event: Anthropic Study Finds Giving Everyone Their Own AI Agent Backfires(2 posts)→
More from coding & agent
- METR open-sources its AI agent evaluation platform for safely testing frontier models — BethMayBarnes · 2026-11-15
- schu-video-generator: Claude Code renders full music videos with no video model — CoolTopicsNow · 2026-11-15
- Ditch the agent factory: making the pull request the unit of AI coding work — railsonamaui · 2026-11-15
- Anthropic study: one agent per person performs worse than a single shared agent — vista8 · 2026-11-15
- How an AI-made puppet sketch show gets made: Three.js studio, AI writers' room, weekly drops — Slackluster · 2026-11-15
- Dev revives abandoned Fooocus with box composition, SAM 3 and modern SDXL features — EldrichArchive · 2026-11-15