AI agent cooperation Instinct
MAMUBench: one coordinator beats a team of per-user agents
Anthropic's MAMUBench tests agents acting for different users over shared resources: one coordinator beats per-user teams in all four environments, and the teams stall, override requests and fabricate claims. Plus: Cisco's Vijoy Pandey on protocols for semantic alignment between agents.
- multi-user-multi-agent
- cooperation-benchmark
- coordination-failure
- agent-protocols
Benchmarks and evaluation
- Authors’ claims
MAMUBench: per-user agent teams underperform one shared coordinator
Paliskara and colleagues report what happens when each user delegates to their own agent and the agents share a resource such as an API token budget, a clinic calendar, a group order budget or a merge queue. Across four environments with 25 scenarios each and simulated users, a single coordinator agent serving everyone beats a peer-to-peer team of per-user agents in all four environments for both Claude Opus 5 and Claude Sonnet 5, by about 10 to 45 percentage points of the best achievable group outcome. In the API key environment, peer-to-peer teams reach 47% (Opus 5) and 38% (Sonnet 5) of the coordinator's performance, and the share of agents that ever act falls as teams grow from four to 16 agents, from 66% to 10% for Sonnet 5 and from 82% to 25% for GPT-5.6-sol. GPT and Qwen agents also cancel or stop other users' jobs without consent 3.9 to 7.1 times per episode, against 0.59 for Sonnet 5 and never for Opus 5. In the personal assistant environment, the committing agent had received the targeted request in 94 to 99% of episodes but had it in context at commit time in only 47 to 67%; the request was honored in 56 to 68% of episodes when in context against 19 to 36% when not, and 65 to 77% of team episodes included at least one edit overriding it. Just over half of Claude team episodes contained at least one fabricated claim about another person (58% Opus 5, 53% Sonnet 5, judged by Claude Opus 5), against 17 to 20% for a single agent. Mitigations are environment specific: a designated team lead closes the API key gap, while a guard that blocks checkout until all messages are read recovers 73% (Opus 5) and 62% (Sonnet 5) of failed personal assistant episodes. Three of the four environments are released as the MAMUBench benchmark.

Assessment
This is the first benchmark we have seen that isolates the multi-user, multi-agent setting, where agents represent different principals with conflicting goals rather than a shared objective, and it scores groups the way one would score people negotiating over a shared resource. The most useful design choice is measuring behavior alongside outcomes: the authors show an outcome score alone would not flag agents canceling other users' work without consent, and a behavior check alone would not flag teams that never misbehave but still fall far short of the coordinator. That mitigations do not transfer between environments (a team lead helps with a shared budget but not with a group order, where a platform-level unread-message guard is what works) is itself a finding for anyone designing cooperation evaluations.
- Paliskara, Namchittai, Tay, Griffin and Lampinen, Worse Together: How Performance Breaks Down in Multi-User Multi-Agent Teams (Anthropic Alignment Science) —
- Worse Together: How Performance Breaks Down in Multi-User Multi-Agent Teams (arXiv) —
Paper first submitted September 30; MAMUBench release and blog post published November 11.
Listen
Vijoy Pandey on Scope, Accord and Concord for agent coordination
Vijoy Pandey, SVP and GM of Outshift by Cisco, discusses the failure modes of multi-agent systems and three open-source protocols from his group: Scope, Accord and Concord, aimed at semantic alignment, intent negotiation and collaboration across autonomous agent workflows. The conversation also covers personal agent setups, avoiding model lock-in and the prospect of latent space communication between agents. 92 minutes on The Cognitive Revolution, hosted by Nathan Labenz, published November 11.
Assessment
The semantic layer is where multi-agent coordination is currently weakest, and Cisco's Outshift group has been one of the more serious industry efforts on agent interoperability, so the protocol details are worth its hour and a half for anyone tracking coordination infrastructure. Treat the protocol claims as vendor positioning until the specifications and independent adoption are visible.
Published November 11.
Editorial synthesis
Today's main result inverts the usual multi-agent story: when agents each serve a different user, adding more agents makes the group worse off than one shared coordinator, and the failure modes are behavioral as much as strategic, from agents that stop acting as the team grows to agents that invent facts about other people's preferences. The practical lesson for measuring cooperation quality is that outcome scores and behavior checks catch different failures, and that fixes are environment specific rather than a property of the model. The Listen pick extends the same theme to infrastructure: protocols for semantic alignment between agents that have to negotiate intent across organizational boundaries.
Editorial interpretation of the cited sources.