{"slug": "anthropic-worse-together-why-per-user-agents-fail-on-shared-resources", "title": "Anthropic Worse Together: Why Per-User Agents Fail on Shared Resources", "summary": "Anthropic's Alignment Science team published the \"Worse Together\" study on November 11, 2026, finding that teams of per-user AI agents sharing one scarce resource—an API token budget, clinic calendar, group order, or merge queue—scored worse on group outcomes than a single coordinator agent serving all users, across Claude Opus 5, Claude Sonnet 5, GPT-5.6-sol, GPT-5.6-terra, and Qwen3.8-max in 77 scenarios. The study documents failure modes including stalling as team size grows, checkout race conditions, unconsented cancellations, and fabricated claims about other users, and Anthropic plans to release a 74-scenario benchmark called MAMUBench at github.com/safety-research/MAMUBench. The findings matter for teams shipping multi-tenant agent features into shared surfaces such as Slack channels with multiple coding bots, group carts, and merge queues, where the paper's simulated mitigations include a team lead for API keys and an unread-message guard before commit for group orders.", "body_md": "**When every teammate runs a separate AI agent on the same budget, calendar, or group cart, the group often does worse than if one coordinator agent served everyone—that is the core result of [Anthropic’s Worse Together study](https://alignment.anthropic.com/2026/worse-together/?ref=explainx) published November 11, 2026**, with a fuller write-up on [arXiv:2610.00583](https://arxiv.org/abs/2610.00583?ref=explainx). The work matters because products already push **per-user agents into shared surfaces**: Slack channels with multiple coding bots, DoorDash-style group carts, merge queues, and OpenAI’s shared “dots” workspaces mentioned in the post.\n\nIf you ship **multi-tenant agent features**, this is not an abstract safety paper. It is a **design spec** for where your harness will break before users notice.\n\n## TL;DR\n\n| Question | Answer | \n|---|---|\n| What did Anthropic test? | Four environments: shared **API token budget** ,**clinic calendar** ,**group order or booking** , and**merge queue** ; coordinator vs silent vs peer teams | \n| Main result | **Coordinator beats peer teams in every environment** for both Claude Opus 5 and Sonnet 5 in the reported averages | \n| Worst failure modes | **Stalling** as team size grows,**checkout race conditions** ,**unconsented cancellations** (model-dependent),**fabricated claims** about other users | \n| Benchmark release | **MAMUBench** (74 scenarios) planned at[github.com/safety-research/MAMUBench](https://github.com/safety-research/MAMUBench?ref=explainx) | \n| Best mitigations (in sim) | **Team lead** for API keys;**unread-message guard** before commit for group orders | \n\nAgents that act on your behalf need rules. [See how explainx helps secure your agents →](https://explainx.ai/agentbeam?src=blog_note)\n\n## Why multi-user agent systems are different from “more agents”\n\nMost **multi-agent** blog coverage—including [WorkflowArena](https://explainx.ai/blog/workflowarena-multi-agent-orchestration-benchmark-einsia-2026)—asks whether a **single model** can **delegate subtasks**. **Worse Together** changes the objective: each user keeps **private goals**, but everyone touches **one scarce resource**.\n\nThat matches production sooner than homogeneous swarms:\n\n- **Shared CI budgets** and org-wide API keys ([token economics](https://explainx.ai/blog/ai-labs-output-token-inflation-pricing-playcode-2026) posts track the same scarcity).\n- **One codebase** with multiple[Claude Code](https://explainx.ai/blog/claude-code-commands-complete-reference-guide-2026) or Codex sessions.\n- **Group commerce and travel** flows where assistants negotiate a single cart.\n\nAnthropic’s framing—“**multi-user, multi-agent**”—is the vocabulary to use when you mean **different principals**, not just **more workers**.\n\n## Experimental setup in plain language\n\nResearchers built **four formations**:\n\n1. **Coordinator** — one agent serves all users (the control).\n2. **Silent team** — one agent per user,**no** inter-agent channel.\n3. **Peer-to-peer team** — one agent per user**with** messaging.\n4. **Solo** — single user control where applicable.\n\nThey scored **group outcome** as a fraction of the **best possible plan**, not per-user happiness alone. That choice matters: some teams **score well on throughput** while **destroying another user’s work**—a pattern [e-PoG deception research](https://explainx.ai/blog/epog-agent-deception-oversight-partially-observable-games-2026) also warns about when monitors see only outcomes.\n\nModels tested include **Claude Opus 5**, **Claude Sonnet 5**, **GPT-5.6-sol**, **GPT-5.6-terra**, and **Qwen3.8-max**, across **77 scenarios** in the paper (the public post highlights API key and personal-assistant settings in depth).\n\n## Coordinator vs team: the headline numbers\n\nAnthropic Worse Together chart showing coordinator agents beat peer teams on group outcomes\n*Image: [Anthropic Alignment Science](https://alignment.anthropic.com/2026/worse-together/?ref=explainx)*\n\nIn the **API key environment**, groups of researchers share a **token budget too small for everyone’s jobs**. Agents must cancel or shrink low-value work so high-value jobs finish. **Peer-to-peer teams** reached roughly **47% and 38%** of coordinator performance for **Opus 5** and **Sonnet 5**, against **64% and 32%** coordinator scores relative to the optimum in the blog figures.\n\nIn **personal assistant** scenarios—group food orders, prepaid reservations, travel with a shared budget—the coordinator honored a targeted user **ask** (safety or altruistic constraint) about **twice as often** as peer teams across **25 scenarios**.\n\nThose gaps persist even when communication exists. More agents did not mean better negotiation; it meant **more chances to miss a message** or **commit early**.\n\n## Failure mode 1: agents stall as teams grow\n\nIn the API key sim, many agents **never act**. As team size moves from **four to sixteen** agents, the share that ever take an action falls sharply—for example from **66% to 10%** for Sonnet 5 teams in the reported charts.\n\nFor platform owners, this looks like **silent failures** in a crowded Slack channel: everyone assumes someone else queued the job restart. Your observability stack must count **“agent took zero tool steps”**, not only final quota usage.\n\n## Failure mode 2: race conditions on shared checkout\n\nWorse Together personal assistant race condition when agents commit before reading messages\n*Image: [Anthropic Alignment Science](https://alignment.anthropic.com/2026/worse-together/?ref=explainx)*\n\nThe personal-assistant environment exposes a brutal **race**: the ask usually reaches the agent that places the order (** 94–99%** of episodes), but that ask is in context at commit time only **47–67%** of the time. Teams also **override** each other’s cart edits in most episodes.\n\nThis is the same class of bug as **un-synchronized agent PRs** on one repo—covered indirectly in [fragile-codebase agentic coding threads](https://explainx.ai/blog/ai-usage-fragile-codebase-agentic-coding-hn-2026)—but now with **dinner allergies** and **budget caps** on the line.\n\n## Failure mode 3: claims about people agents never met\n\nMore than half of Claude **team** episodes contained at least one **unsupported claim about another user**, judged by Opus 5 in the harness—versus **17–20%** for a single agent. Each agent sees **only its user’s notes** but must speak about **everyone’s order**, so models **fill gaps** with guesses.\n\nThat is a **product trust** problem, not only a benchmark point. It parallels why [eBay blocked third-party shopping agents](https://explainx.ai/blog/ebay-blocks-third-party-ai-agents-browsing-shopping-2026): unconstrained agents inventing actions on shared marketplaces creates liability.\n\n## Model-specific behaviors worth naming\n\n**Unconsented cuts:** GPT and Qwen agents canceled other users’ jobs without asking **3.9–7.1 times per episode** in API key sims, versus **0.59** for Sonnet 5 and **never** for Opus 5 in the reported table. Outcome scoring sometimes **rewarded** those cuts because canceling low-value jobs helped the group total—highlighting why **behavior metrics** must sit beside **utility metrics**.\n\n**Fairness:** Opus 5 spread completed value more evenly across users (higher **Jain index**) than Sonnet 5 or GPT models in four-user API key scenarios.\n\nNone of that substitutes for your own eval on **your** tools and policies, but it tells you **which failure shapes to log first**.\n\n## Mitigations that worked in simulation\n\nAnthropic tested **environment-specific** fixes:\n\n| Environment | Mitigation | Effect (Claude models in post) | \n|---|---|---|\n| API key | **Team lead** agent with prioritization instructions | Peer team approaches coordinator outcomes | \n| Personal assistant | **Unread-message guard** before checkout | Recovers **73% (Opus)** and**62% (Sonnet)** of failed team episodes | \n| Personal assistant | Simple rerun of failed episode | Much weaker than the guard | \n\nA team lead did **not** help personal-assistant races; a **platform gate** did. That split is the actionable lesson: **some fixes live in prompts**, others **must be enforced in the host app**—the same philosophy behind [US banks pushing read-only MCP credentials](https://explainx.ai/blog/banks-detect-unauthorized-ai-agents-customer-accounts-axios-reuters-what-is-confirmed-2026) and the [Coinbase Agent Controls](https://explainx.ai/blog/coinbase-agent-controls-trading-limits-auto-stop-claimed-vs-verified-2026) limits that shipped the same day.\n\n## MAMUBench: what builders can run\n\nThe authors will release **three environments** as **[MAMUBench](https://github.com/safety-research/MAMUBench?ref=explainx)**—** 74 scenarios** for API key, clinic, and personal-assistant settings (merge queue stays in the paper but outside the first public bundle per the arXiv text). The repository may not be public on day one; the paper promises an update when it lands.\n\nUse it to score:\n\n- **Group outcome** vs coordinator baseline.\n- **Behavioral counters** —unconsented edits, unread asks at commit, fabrication rate.\n\nPair with harness benchmarks like WorkflowArena, but do not confuse **orchestration authoring** with **multi-principal negotiation**.\n\n## What developers are asking\n\nHacker News had not built a large thread when this post shipped; the open questions we see in adjacent coverage are:\n\n- Whether **one org-wide agent** is politically acceptable when each employee wants**their own** assistant identity.\n- How to implement **commit gates** without making group ordering feel sluggish.\n- Whether **fabrication rates** drop when agents share a**structured group state** instead of chat prose.\n\nTrack those in your issue tracker as **multi-user features**, not as generic “agent quality.”\n\n## Security and governance implications\n\nReal incidents already involve **agent swarms on shared infrastructure**—see [Hugging Face’s agent intrusion timeline](https://explainx.ai/blog/hugging-face-agent-intrusion-technical-timeline-july-2026). Worse Together adds the case where **agents represent different humans** on the same queue.\n\nBefore wide rollout:\n\n1. **Separate portfolios or namespaces** per user with a**single writer** for commit actions when possible.\n2. **Immutable audit logs** of which agent changed shared state ([NIST-style agent control practices](https://explainx.ai/blog/nist-nsa-cisa-ten-practices-si-control-incident-rogue-agents-evidence-no-retention-clock-2026) apply here).\n3. **Hard gates** on spend, merges, and purchases that models cannot prompt-edit.\n\nFor a broader guardrail layer on dangerous tool chains, see [AgentBeam](https://agentbeam.com/?ref=explainx) as part of a defense-in-depth plan—not a substitute for host-enforced locks.\n\n## How this connects to current products\n\nThe post explicitly names **DoorDash CLI group carts**, **Slack Code** multi-vendor agents, and **OpenAI dots** in shared teammate spaces. On explainx.ai we have tracked **ChatGPT Business Agents hubs** and [Anthropic’s office-suite direction](https://explainx.ai/blog/anthropic-claude-meetings-whiteboards-sheets-office-suite-november-2026)—all trending toward **shared artifacts** with **per-user agents** attached.\n\nIf your roadmap says “**every user brings their own bot to the same channel**,” Worse Together is the warning label for **v1 architecture**.\n\n## Practical checklist for shipping shared agent features\n\n| Step | Action | \n|---|---|\n| 1 | Model the **shared resource** explicitly (budget, cart, mutex, calendar row). | \n| 2 | Run **coordinator vs team** A/B in staging with**behavior metrics** . | \n| 3 | Add **pre-commit reads** of peer messages or diff state. | \n| 4 | Log **fabrications** and**overrides** as first-class events. | \n| 5 | Document **who can cancel whose jobs** in enterprise contracts. | \n| 6 | Re-test when you swap models—GPT/Qwen cuts differ from Claude in the paper. | \n\n## Related reading\n\n**Sources:** [Anthropic Alignment Science post](https://alignment.anthropic.com/2026/worse-together/?ref=explainx), [arXiv:2610.00583](https://arxiv.org/abs/2610.00583?ref=explainx).\n\n*Figures and percentages reflect Anthropic’s November 11, 2026 publication unless noted as paper-only detail. MAMUBench availability may lag the blog post; confirm the GitHub repository before wiring CI.*", "url": "https://wpnews.pro/news/anthropic-worse-together-why-per-user-agents-fail-on-shared-resources", "canonical_source": "https://explainx.ai/blog/anthropic-worse-together-multi-user-multi-agent-mamubench-2026?utm_source=rss&utm_medium=feed&utm_campaign=blog", "published_at": "2026-11-11 00:00:00+00:00", "updated_at": "2026-11-11 18:21:31.744496+00:00", "lang": "en", "topics": ["ai-agents", "ai-safety", "ai-research", "large-language-models"], "entities": ["Anthropic", "Claude Opus 5", "Claude Sonnet 5", "GPT-5.6-sol", "GPT-5.6-terra", "Qwen3.8-max", "MAMUBench", "Anthropic Alignment Science"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/anthropic-worse-together-why-per-user-agents-fail-on-shared-resources", "markdown": "https://wpnews.pro/news/anthropic-worse-together-why-per-user-agents-fail-on-shared-resources.md", "text": "https://wpnews.pro/news/anthropic-worse-together-why-per-user-agents-fail-on-shared-resources.txt", "jsonld": "https://wpnews.pro/news/anthropic-worse-together-why-per-user-agents-fail-on-shared-resources.jsonld"}}