Anthropic releases MAMUBench: multi-agent teams perform worse together on shared resources
AndrewLampinen · x · 2026-11-12
Anthropic Fellows research "Worse Together" (Sahan Paliskara, Andrew Lampinen et al.) releases MAMUBench, a benchmark for multi-user multi-agent coordination.
Setup: four simulated everyday environments where agents share resources (compute budgets, codebases, group orders), testing frontier models from Anthropic, OpenAI and Alibaba.
Findings:
- In all environments, one-agent-per-user teams did worse than a single coordinator serving everyone; silent teams without peer channels collapsed (0/3 tasks) in two environments.
- Peer-to-peer teams became too cautious, forgot tasks, overrode peers' requests, and made unsupported claims about others, varying by model and environment.
- Fixes were environment-specific: designated team leader + stated priorities helped shared budgets; blocking orders until messages were read rescued more failed group orders than rerunning.
Three environments are open-sourced as MAMUBench with paper and code.
Related event: Anthropic Study Finds Multi-User AI Agents Perform "Worse Together"(3 posts)→
More from coding & agent
- GitHubToDo: open-source native Windows GitHub client with Copilot AI summaries — unixterminal · 2026-11-12
- One line in AGENTS.md took coding agent pass rate from 16% to 100% — rseroter · 2026-11-12
- Coded cartoon show made with Fable 5.2 hints at interactive LLM entertainment — mickeyxfriedman · 2026-11-12
- Dev praises tldraw offline + Omarchy as a rapid prototyping combo — SubtleGradient · 2026-11-12
- Dev rebuilds open-source desk robot Robbi with PCB designed almost entirely by Codex — OpenAIDevs · 2026-11-12
- Mock everything, keep control: a developer's real job in the AI coding era — 0xkarasy · 2026-11-12