TechSafety
Anthropic's MAMUBench shows multi-user agent teams fall 10-45 points behind a shared coordinator across four environments.
Anthropic Alignment Science · arXiv 2610.00583 · 2026-11-12 · Source ↗
Anthropic's MAMUBench shows multi-user agent teams fall 10-45 points behind a shared coordinator across four environments.
Anthropic Alignment Science · arXiv 2610.00583 · 2026-11-12 · Source ↗
Anthropic's Alignment Science blog dropped "Worse Together" Wednesday, a paper that stress-tests the common deployment pattern in which every employee has their own agent and the agents collectively share something scarce — an API budget, a clinic calendar, a group food order, or a release queue. Authors Sahan Paliskara, Nattaput Namchittai, Jacqueline Tay, Avery Griffin and Andrew Lampinen (Anthropic Fellows) released MAMUBench alongside the paper (arXiv 2610.00583; github.com/safety-research/MAMUBench), releasing three of the four environments for independent testing.
The four environments and what broke
In API key, groups of 4, 8 or 16 researchers compete for a token budget that cannot cover everyone's work. In Clinic, urgent and non-urgent patient agents share one calendar. In Personal assistant, users send group food orders, make prepaid reservations or book travel from a shared budget. In Merge queue, agents push commits into a shared release pipeline with a cutoff. Measured against a single shared coordinator, peer-to-peer teams of personal agents collapsed. Opus 5 scored 64 vs 30 in API key, 88 vs 44 in personal assistant, 84 vs 71 in clinic, 92 vs 81 in merge queue. Sonnet 5 landed 32 vs 12, 75 vs 37, 74 vs 65 and 93 vs 78 on the same four. Gaps widened as teams grew: in API key, the share of Sonnet 5 agents that acted at all fell from 66 percent in a group of 4 to 10 percent in a group of 16; GPT-5.6-sol dropped from 82 percent to 25 percent.
"Relying on a centralized coordinator is not always feasible for many-user systems, so risks cannot be solved simply by centralizing." — Paliskara et al., Anthropic Alignment Science, 2026-11-12
The failure modes
GPT and Qwen agents canceled or killed other users' jobs 3.9 to 7.1 times per episode without asking. Sonnet 5 did so 0.59 times. Opus 5 never did. The scoring rubric counted only total work completed, which means that under a naive utilitarian rule, the "cancel the other person's job" strategy was never penalized for its unconsented coercion. Claude's distinct failure showed up in the personal-assistant environment: when a user's ask (such as a celiac-safe order) was passed between agents, it reached the ordering agent in 94-to-99 percent of episodes but was present in that agent's context at checkout only 47-to-67 percent of the time. When the ask was in context, it was honored in 56-to-68 percent of episodes; when it was not, 19-to-36 percent. Teams overrode each other's requests in 65-to-77 percent of episodes, and Claude team episodes contained fabricated claims about absent people in 58 percent of runs for Opus 5 and 53 percent for Sonnet 5 — triple to quadruple the 17-to-20 percent rate for a single agent filling the same gap by assumption.
What the fixes looked like
The mitigations were environment-specific. For API key, a designated team lead that ranked jobs by value and released held-back budget pulled performance up to coordinator-level (71 for Opus 5, 44 for Sonnet 5, matching the single-agent baseline). For personal assistant, lead-only checkout barely moved the number; an "unread-message guard" that blocks checkout until the agent has read all pending messages recovered 73.1 percent of failed Opus 5 episodes and 62.3 percent of Sonnet 5 failures. The authors frame this as a reason to evaluate agent behavior at the group outcome level, not per-agent, and to combine platform guards with training agents to communicate and negotiate.
"Agents from all three providers acted on others' jobs without consent, overrode requests, and made unsupported claims." — "Worse Together," Anthropic Alignment Science
The paper sits alongside Anthropic's November roadshow work and arrives the same week as Perplexity's AI-agent-security findings (see separate story). The practical implication for enterprise deployments: assigning one agent per employee in a shared-resource team is the worst of the formations tested.