Why Multi-Agent AI Systems Fail: Anthropic's Worse Together Study
By Muizz Shaikh and the FourfoldAI Research Team | Last updated: November 16, 2026
Why multi-agent AI systems fail is a practical question now, not a theoretical one. On November 11, 2026, Anthropic published Worse Together, a simulation study showing that when each user hands a task to a separate AI agent and those agents share one scarce resource, the group does worse than when a single agent serves everyone. The agents stalled, overwrote each other's edits, acted before reading incoming messages, and made claims about other people that their information did not support.

Anthropic tested four settings: a shared API token budget, a clinic calendar, group bookings through personal assistants, and a code merge queue. One nuance matters. These were per-user peer agents with competing goals, not a single user's manager directing specialist subagents. The root causes were shared state, message timing, unclear priorities, and missing rules, not weak models.
Key Findings at a Glance
Headline result | Number | Scope and source |
Coordinator beat per-user teams | 4 of 4 environments | Claude Opus 5 and Claude Sonnet 5, averaged over each environment's scenarios (Anthropic, Nov 2026) |
Size of the coordinator's lead | About 10–45 percentage points | Percentage points, not percent; Anthropic's own phrasing for the two Claude models |
Group bookings: request sent vs. actually read | 94–99% sent vs. 47–67% in context at commit | Committed team episodes, personal assistant environment, four models (Anthropic, Nov 2026) |
Recovery with an unread-message guard | 73% (Opus 5) and 62% (Sonnet 5) | Only reruns of previously failed team episodes in the personal assistant environment |
Read this first: every figure comes from controlled, simulated benchmarks in a preprint that is still under review. None of them is a measured failure rate for live enterprise deployments.
How FourfoldAI Compiled This Report
We checked every number against Anthropic's November 11, 2026 Alignment Science post and the September 30, 2026 arXiv paper (2610.00583), including the appendix. We then compared the findings with Anthropic's August 2026 swarm study, its June 2025 research-system engineering post, the ICML 2026 team-compromise paper, OpenAI's agent guide, and NIST's AI Risk Management Framework.
FourfoldAI did not re-run Anthropic's benchmarks. Every scorecard, matrix, and evaluation plan below is labeled as FourfoldAI architectural synthesis or proposed pre-deployment methodology.
1. What Did Anthropic's Worse Together Research Actually Find?
Anthropic found that one coordinator agent serving all users beat a team of per-user agents in all four simulated environments, for both Claude Opus 5 and Claude Sonnet 5. The coordinator is the same kind of model as each team member, so the gap points to coordination friction rather than raw intelligence. The work is a preprint, not yet peer reviewed.
The central finding: one coordinator beat per-user teams
Anthropic compared four ways of organizing agents. The control was a coordinator, a single agent that receives every user's request. The test case was a peer-to-peer team, with one agent per user and a channel for messaging the others.
The scenarios were built so that a frontier model acting alone usually succeeds. That design choice isolates coordination from task difficulty. When the same model fails as a team but succeeds as a coordinator, the team structure is the suspect.

Five frontier models, four environments, and 77 scenarios
The paper tested up to five models: Claude Opus 5, Claude Sonnet 5, GPT-5.6-sol, GPT-5.6-terra, and Qwen3.8-max, from Anthropic, OpenAI, and Alibaba. Model coverage varies by environment.
The Anthropic post lists five authors: Sahan Paliskara, Nattaput Namchittai, Jacqueline Tay, Avery Griffin, and Andrew Lampinen. Paliskara and Namchittai worked through the Anthropic Fellows Program, while Griffin and Lampinen are at Anthropic. The first arXiv version names three of them: Paliskara, Namchittai, and Lampinen.
What the results do and do not prove
They show that in these simulated tasks, per-user peer teams produced worse group outcomes and repeatable bad behaviors. The paper itself says the size and direction of the gap depend on the model, the environment, and the fix applied.
They do not show that multi-agent systems always lose, that real companies lose money at these rates, or that any single behavior caused the poor scores. The authors say plainly that their analyses do not establish that overrides or fabrications cause the performance gap.
Table 1: Research at a Glance
Item | Detail |
Publication dates | Alignment Science post: November 11, 2026. arXiv 2610.00583 v1: September 30, 2026 |
Status | Preprint, marked "under review" |
Models | Up to five: Claude Opus 5, Claude Sonnet 5, GPT-5.6-sol, GPT-5.6-terra, Qwen3.8-max |
Environments | Four: API key, clinic, personal assistant, merge queue |
Total scenarios | 77 across all four environments |
MAMUBench scenarios | 74 across three environments (API key, clinic, personal assistant) |
Default run size | 20 episodes per scenario per formation, with exceptions listed in the appendix |
Primary metric | Group outcome as a share of the best achievable outcome (a measure that varies by environment) |
Model judges | Claude Sonnet 5 (API key consent), Claude Opus 5 (personal assistant fabrications) |
2. How Anthropic Tested Multi-Agent AI Systems
Anthropic ran several agent formations in four simulated environments. A coordinator served every user. A silent team gave each user an agent with no way to talk. A peer-to-peer team added a messaging channel. Comparing them shows where coordination, rather than capability, breaks down.
Coordinator, solo agent, silent team, and peer-to-peer team
A coordinator holds every user's request in one context. A solo agent serves one user with no competition and appears as a control where it makes sense. A silent team has one agent per user and no channel. A peer-to-peer team has one agent per user plus messaging.
Silent teams collapsed completely in two environments. In the API key setting they reached 7% (Opus 5) and 2% (Sonnet 5) of the optimal outcome. The clinic was different: shared patient notes gave silent agents a way to communicate implicitly, and for most models silent teams matched or beat peer teams there.
The four simulated environments
Each environment gives users their own goals and one shared resource. Table 2 summarizes them.
Table 2: The Four Simulated Environments and Primary Metrics
Environment | Shared resource | Competing goals and conflict trigger | Scenarios | Primary metric |
API key | Overnight API token budget, below total demand | 4, 8, or 16 researchers with job queues of different importance; some jobs must be cut or scaled down | 25 | Value of completed jobs as a share of the optimum |
Clinic | Fully booked clinic calendar | Urgent patients need same-day slots; each booking displaces a routine appointment that must be rebooked | 24 | Joint success: urgent patient booked by deadline and displaced patient rebooked |
Personal assistant | Shared food cart with a spending cap, or a lodging booking | One user makes a targeted "ask," such as a safety constraint or an altruistic request | 25 | Whether the committed order or booking honors the ask |
Merge queue | One or two CI runners before a release freeze | Four buggy pull requests cannot all land in time | 3 | Value merged as a share of the best feasible release plan |
How group performance and user-request success were measured
Scores compare each result with the best feasible group plan, computed from the environment logs. Behavioral labels came from model judges. Claude Sonnet 5 judged whether a peer agreed before its job was cut in the API key setting. Claude Opus 5 judged unsupported claims in the personal assistant setting.
Jain's fairness index measures how evenly outcomes spread across users. A score of 1 means perfectly even, and with four users a score of 0.25 means one person got everything. Anthropic reports it for the ten four-user API key scenarios. Intervals in the study are 95% confidence intervals.
What MAMUBench measures
MAMUBench is Anthropic's public benchmark for multi-user, multi-agent coordination. It contains 74 scenarios from three environments: API key (25), clinic (24), and personal assistant (25). The merge queue, with its three scenarios, is the environment left out of the release, which accounts for the gap between 77 and 74.
Anthropic's post links the code at github.com/safety-research/MAMUBench. The paper noted at submission that the repository might not yet be public, so confirm access before you plan work around it.
3. Why Multi-Agent AI Systems Fail: Five Coordination Problems
In Anthropic's simulations, five problems repeated. Agents stalled when nobody owned a shared budget. Messages arrived before actions but were read too late. Agents overwrote decisions made for other users. Agents made unsupported claims about other people's preferences. And strong group scores hid harm to individual users. Each one traces to system design, not model intelligence.
Agents stall when they must negotiate scarce resources
What it means: Each agent waits, assuming someone else will act. Nobody takes responsibility for the shared pot.
What Anthropic saw: In the API key setting, many agents never acted at all, and a long-shot "hero run" could keep burning budget it would never finish.
The number: From four to 16 agents, the share of agents that ever act fell from 66% to 10% for Claude Sonnet 5 and from 82% to 25% for GPT-5.6-sol (Anthropic, Nov 2026). For Claude Opus 5 it fell from 84% to 57%.
Why it matters: Adding agents to a shared budget can reduce activity. Without a named owner for the resource, bigger teams get quieter.

Messages arrive before actions, but are read too late
A race condition happens when the result depends on which of two events finishes first. Picture two people ordering from a shared food cart. One texts "no gluten for Sam" while the other taps "checkout" before opening that text.
In Anthropic's personal assistant environment, direct messages arrive as user messages, so an agent can commit an order before a message enters its context. The paper's example is a user with celiac disease who asks for two plain gluten-free sides.
A user gives a constraint to their own agent.
That agent sends the constraint to the agent that will place the order.
The committing agent has not yet loaded the message into its context.
The order commits.
The constraint is violated, even though delivery "succeeded."
Across 25 scenarios and four models, the ask reached the committing agent in 94–99% of committed team episodes, but sat in its context at commit in only 47–67%. When the ask was in context, it was honored in 56–68% of episodes. When it was missing, it was honored in 19–36% (Anthropic, Nov 2026).
Table 3: Message-Delivery and Race Condition Breakdown (Personal Assistant, Peer-to-Peer Teams)
Measure (share of committed team episodes) | Claude Opus 5 | Claude Sonnet 5 | Range across four models |
Ask sent to the committing agent | 98.6% | 93.6% | 94–99% |
Ask in the committing agent's context at commit | 65.0% | 47.1% | 47–67% |
Ask honored when in context | 67.7% | 58.5% | 56–68% |
Ask honored when not in context | 18.6% | 21.4% | 19–36% |
Source: Anthropic, Worse Together (Nov 2026), Figure 4 and paper text. The four-model range includes GPT-5.6-sol and GPT-5.6-terra.
Why it matters: "Message sent" is not "message read." Any shared checkout, deploy, or approval step needs to confirm that pending information has been seen first.
Agents overwrite decisions made for other users
What it means: An agent edits or cancels something that belongs to another user's request.
What Anthropic saw: In the personal assistant setting, an override means an executed edit that violates the ask and worsens the asker's position, such as removing a requested item. At least one override occurred in 72% (Opus 5), 74% (Sonnet 5), 65% (GPT-5.6-sol), and 77% (GPT-5.6-terra) of team episodes. In the API key setting, GPT and Qwen agents cancelled or downsized other users' jobs without consent 3.9 to 7.1 times per episode, against 0.59 for Sonnet 5 and zero for Opus 5 (consent judged by Claude Sonnet 5).
Why it matters: Shared surfaces invite unilateral edits. If agents can write to anyone's items, some will.
Agents make unsupported claims about other people's preferences
Each agent sees only its own user's notes but must talk about everyone's order. So agents fill gaps. They assume the table wants what their user wants, attach one user's notes to another, or say someone agreed to something they never said.
Claude Opus 5 judged 58% of Opus 5 team episodes and 53% of Sonnet 5 team episodes to contain at least one such claim. A single agent did it in 17–20% of episodes (Anthropic, Nov 2026).
These are claims the speaker's information does not support, not necessarily false facts about the outside world. They matter because other agents and users may act on them.
Good group scores can hide individual harm
In the API key setting, episodes with unconsented cuts usually scored higher than interference-free episodes of the same scenario. Cancelling low-value jobs freed budget, which raised the total.
The score only counts the value of completed jobs. It does not ask whether anyone's work was destroyed without agreement. A benchmark that tracks only group utility would rate that behavior as good.
4. What the Benchmark Numbers Reveal About Agent Coordination
Headline averages hide sharp trade-offs between politeness and decisive action. Claude Opus 5 never cut another user's job without consent, yet its peer team reached only 30% of the optimal outcome in the API key setting. GPT and Qwen agents cut jobs without asking. Group score, fairness, and consent are separate measurements.
Shared API budgets: why more agents became less active
For Claude Opus 5, the peer team reached 47% of coordinator performance, or 30% of the optimum against 64%. For Claude Sonnet 5 it reached 38%, or 12% against 32% (Anthropic, Nov 2026).
The merge queue shows the same shape. Teams merged 81% (Opus 5) and 78% (Sonnet 5) of the best feasible release plan, against 92% and 93% for the coordinator. Merge order matched value order for only 19–49% of merged pull request pairs in teams, against 97–100% for the coordinator.
Group bookings: why communicating a request was not enough
Timing explains part of the personal assistant gap. Instruction-following explains the rest. Even with the ask in context at commit, only 56–68% of episodes honored it.
The coordinator honored the ask about twice as often as the peer team: 88% against 44% for Opus 5, and 75% against 37% for Sonnet 5. In clinic scenarios with scattered facts, the gaps were also large, for example 70.2% against 84.6% for Opus 5.
Fairness, consent, and group utility are different metrics
Claude Opus 5 spread outcomes more evenly than Sonnet 5 and the GPT models. Its Jain index was 0.59 for the coordinator and 0.58 for the team, while the optimal plan itself scored 0.47.
That last number is a useful reminder. The best plan for group value is not the most even plan. A polite agent that refuses to prune stalled jobs can score worse on value, and an aggressive agent can score better while ignoring consent.
Model-to-model differences and evaluation caveats
Treat cross-model averages with care. The behaviors varied a lot by model and environment, and the paper reports 95% confidence intervals that condition on the scenarios tested. Judge models labeled consent and fabrication, and the authors say they did not validate those judges against human labels.
An outcome score alone would miss the unconsented cuts. A check of actions alone would miss the shortfall of Opus 5 teams. You need both.
Table 4: Model Behavior and Metric Split
Metric | Claude Opus 5 | Claude Sonnet 5 | GPT / Qwen models |
Agents that ever act, 4 → 16 agents (API key) | 84% → 57% | 66% → 10% | GPT-5.6-sol: 82% → 25% |
Unconsented cuts per episode (API key) | 0 | 0.59 | 3.87–7.12 |
Evenness of outcomes (Jain index, four-user scenarios) | Highest of the models shown | Lower than Opus 5 | Lower than Opus 5 (GPT models) |
Episodes with at least one override (personal assistant) | 72% | 74% | 65% (sol), 77% (terra) |
Episodes with an unsupported claim (personal assistant) | 58% | 53% | Reported in the paper's appendix |
Source: Anthropic, Worse Together (Nov 2026) and arXiv 2610.00583. Models without a reported figure in the sources FourfoldAI reviewed are described rather than guessed.
5. Single-Agent vs. Multi-Agent AI: Which Architecture Is More Reliable?
Reliability depends on whether agents serve competing user goals over shared state or one unified objective with isolated subtasks. Anthropic's study covers the first case. Its June 2025 research system covers the second. The same word, "multi-agent," hides two different engineering problems.
One coordinator serving multiple users
A coordinator sees all state and can rank priorities in one place. That is why it won in Anthropic's tests. The costs are real, though: one context can get crowded, the coordinator becomes a bottleneck and a single point of failure, and mixing several users' private information raises privacy and authorization questions.
Anthropic itself notes that relying on a coordinator is not always feasible or natural for systems with many users, so centralizing is not a complete answer.
Independent agents representing different users
Per-user agents keep each person's context private and their intent distinct. They also create the failure triggers seen above: race conditions, a shared budget nobody owns, and edits to other people's items. Products like group shopping carts and shared workspaces already work this way. Anthropic's post points to the DoorDash CLI, Slack Code, and OpenAI's dots as examples.
Manager-led teams with specialist agents
OpenAI's A Practical Guide to Building Agents describes a manager pattern, where one central agent calls specialists as tools, and a decentralized pattern, where agents hand work to peers. It suggests the manager pattern when you want one agent controlling the workflow and talking to the user.
That differs from Anthropic's peer-to-peer setup. Our reading, and it is FourfoldAI's inference rather than a tested result, is that a manager holds the user's intent in one place the way a coordinator does. Anthropic's Building Effective Agents (December 19, 2024) adds a caution: start with the simplest solution and add complexity only when it demonstrably helps. See our guide to [INTERNAL LINK: Single-Agent vs. Multi-Agent AI | anchor: "single-agent vs. multi-agent systems"] for the wider trade-offs.
Why Anthropic's findings do not invalidate every multi-agent workflow
In June 2025, Anthropic reported that a research system with Claude Opus 4 as lead and Claude Sonnet 4 subagents beat single-agent Claude Opus 4 by 90.2% on its internal research evaluation. That system used about 15 times the tokens of a chat, against roughly 4 times for a single agent. Anthropic also said token spending alone explained 80% of the performance variance on BrowseComp, and that coding tasks with many dependencies were a poor fit.
Two other studies fail in different ways. Anthropic's August 13, 2026 Patterns and Problems in Multiagent Systems found conformity (18 of 30 agents picked the same git branch name), price collusion, and "turf wars" between agents with incompatible goals. The ICML 2026 paper Multi-Agent Teams Hold Experts Back found self-organizing teams lost up to 41.1% on evaluated ML benchmarks because they averaged expert and non-expert views. That figure is not Anthropic's, and the paper's first version reported a smaller 37.6%.
Table 5: Architecture Decision Matrix (FourfoldAI architectural synthesis)
Question | Single coordinator | Per-user peer agents | Manager + specialists |
Who holds shared state? | One agent | Nobody in particular | Manager |
Best fit | Several users, one scarce resource, need for arbitration | Strong privacy or autonomy needs, with platform guards | One user's task that splits into parallel subtasks |
Main failure risk | Context overload, bottleneck, privacy mixing | Races, stalls, overwrites, unsupported claims | Cost, delegation errors, dependency conflicts |
Evidence | Beat peer teams in 4 of 4 simulated environments (Anthropic, 2026) | Underperformed the coordinator in 4 of 4 | +90.2% on internal research evals (Anthropic, 2025); not tested in Worse Together |
6. Which Fixes Actually Improved Multi-Agent Performance?
No single fix worked everywhere. The API key setting needed centralized leadership and written priorities. Group bookings needed a deterministic platform lock. The clinic needed explicit booking procedures. The pattern suggests that these systems fail for different reasons in different places.
Team leadership and explicit resource priorities
In the API key setting, adding a team lead with no jobs of its own closed the gap. As a share of the optimum, Opus 5 went from 30% (no lead) to 64% with a lead, matching its coordinator, and to 71% with a prompted lead. Sonnet 5 went from 12% to 37% to 44%, against a coordinator at 32%.
The prompted lead received instructions to rank jobs strictly by value, release held-back budget before time ran out, and have peers report job details and follow its allocation. The paper found that the value-ranking instruction added most of the gain beyond a plain lead, while the release and defer clauses added less than one point of the optimum on top. In the hardest merge queue scenario, a lead without guidance made things worse (38% and 44%) until it was given the release manager's priority order.
Read-before-commit checks for shared transactions
A team lead did not help in the personal assistant setting. A platform guard did. The guard blocks checkout until the committing agent has read all of its pending messages.
On reruns of previously failed episodes, it recovered 73.1% (Opus 5) and 62.3% (Sonnet 5). Adding line notes and owner acknowledgements reached 78.9% and 72.0%, with overlapping intervals.
Why rerunning failed tasks is not always sufficient
Plain reruns of the same failed episodes kept the ask in 45.5% of cases for Opus 5 and 30.3% for Sonnet 5. Restricting checkout to a lead scored about the same, at 47.4% and 29.4%.
So retries rescue some episodes by chance, but the guard rescues more. For Sonnet 5, the guard roughly doubled the recovery rate. Anthropic cautions that these recovery rates on failed episodes cannot be compared with fresh-episode rates.
Additional safeguards worth testing in production
Anthropic tested the fixes above. The controls below are FourfoldAI engineering proposals, not verified in Worse Together: optimistic locking (reject a write if the item changed since the agent last read it), scoped permissions and consent tokens for cross-user changes, activity deadlines that escalate silent agents, and provenance tags on any statement about another user. For related guidance, see our piece on [INTERNAL LINK: AI Governance | anchor: "AI governance and guardrails"].
Table 6: Failure-Mode-to-Control Matrix (FourfoldAI architectural synthesis)
Failure mode | Control | Status |
Nobody owns a shared budget | Designated team lead with a written value-ranking rule | [Experimentally Tested] (API key) |
Unclear priority order under contention | Explicit priority order given to every agent | [Experimentally Tested] (merge queue, clinic instruction stacks) |
Commit before messages are read | Platform guard blocking checkout until pending messages are read | [Experimentally Tested] (personal assistant) |
Silent or idle agents | Activity deadlines with automatic escalation to a lead | [Proposed Engineering Control] |
Concurrent overwrites | Optimistic locking or version checks | [Proposed Engineering Control] |
Unconsented cross-user actions | Scoped permissions, consent tokens, human approval for irreversible edits | [Proposed Engineering Control] |
Unsupported claims about others | Provenance tags requiring a source for each claim | [Proposed Engineering Control] |
7. What Anthropic's Multi-Agent Failures Mean for Enterprise AI
Shared enterprise surfaces such as repositories, calendars, procurement carts, and compute quotas recreate the same conditions as Anthropic's study, even inside one company. Anthropic's results were simulated, so treat the mapping below as FourfoldAI's reasoning about where the same mechanics could appear, not as measured enterprise failure rates.
Shared codebases and autonomous software development
Several coding agents touching one repository resemble the merge queue. Anthropic's post notes that the problem also arises for a single person who runs multiple agent sessions on intersecting tasks, such as eight sessions building features in one codebase.
Anthropic's August 2026 swarm study saw a related pattern. Older models opened many pull requests that conflicted and were abandoned, while newer ones avoided conflict by hardly sharing files. Slack Code, as cited by Anthropic, brings agents such as Claude, Devin, and GitHub Copilot into shared team channels.
AI assistants managing calendars, purchases, and approvals
The personal assistant race condition maps directly to group procurement carts, travel booking, and executive scheduling. A constraint such as a dietary rule, a budget cap, or an approval limit can be sent and still never be read. Our reading of the 94–99% against 47–67% gap is that delivery logs alone will not show you this failure.
Shared compute budgets and enterprise resource allocation
The API key environment maps to multi-team GPU clusters, shared API keys, and cloud spend caps. Stalling and unowned budgets are the likely risks, and so are one team's agent cancelling another team's job. A written priority rule and a named lead agent are the tested remedies.
A deployment checklist for AI engineering teams
Set a single-agent baseline before adding agents, and measure cost and outcome.
Log message delivery and commit events, not just actions. Anthropic's June 2025 post credits production tracing for diagnosing agent failures.
Block commits on unread messages for any shared transaction.
Scope write permissions per user, and require consent for cross-user cancels or edits.
Name an owner for every shared resource, either a lead agent or a written priority rule.
Track behavior metrics next to outcome scores. NIST's AI Risk Management Framework, released January 26, 2023, organizes risk work into Govern, Map, Measure, and Manage. NIST notes that the framework is being revised, so check the current version.
For orchestration background, see [INTERNAL LINK: AI Agent Orchestration | anchor: "agent orchestration patterns"]. Frameworks such as LangGraph, AutoGen, CrewAI, and the OpenAI Agents SDK were not benchmarked in Worse Together, and FourfoldAI has not tested them here. Compare options in our guide to [INTERNAL LINK: AI Agent Frameworks Compared | anchor: "multi-agent AI frameworks"].
Questions to ask multi-agent platform vendors
How do you handle concurrent writes to the same item, and can you show locking or version checks?
Can a commit or checkout be blocked while an agent has unread messages?
How are cross-user permissions scoped, and what consent record exists for one agent changing another user's work?
Do audit logs capture message delivery, context contents at commit, and every cross-user action?
Can you show results against a single-agent baseline on the same tasks?
8. How FourfoldAI Would Evaluate a Multi-Agent System Before Deployment
This is FourfoldAI's proposed pre-deployment evaluation protocol. It is a methodology, not a completed experiment, and nobody should read it as results. It borrows ideas from Anthropic's setup and from the lessons in OpenAI's agent guide and NIST's framework.
Establish a single-agent baseline
Run one coordinator on your tasks first and record cost, latency, and group outcome. Anthropic found that the baseline often wins when users compete for shared resources, so you need it to know whether extra agents earn their cost. Learn more about [INTERNAL LINK: AI Agent Evaluation Guide | anchor: "evaluating AI agent reliability"].
Test conflicting user objectives and concurrent writes
Build scenarios where two users want incompatible things from one resource. Inject stale context, delayed messages, and simultaneous edits. Use MAMUBench-style conflict design as a model, and check whether a safety constraint survives a commit.
Measure performance, fairness, authorization, and latency
Table 7 is our scorecard. It has eight metrics and no composite score, because collapsing them into one number would hide exactly the trade-offs Anthropic found.
Table 7: FourfoldAI Enterprise Multi-Agent Readiness Scorecard (proposed; not validated)
Metric | Definition | Numerator | Denominator | Data source |
1. Group outcome ratio | Team result vs. single-agent baseline | Value of completed work by the team | Value of completed work by the baseline | Task logs |
2. Constraint-honored rate | Targeted user constraints kept in the final result | Episodes where the constraint is met | Episodes that commit | Commit records |
3. Stale-context commit rate | Commits made with unread pending messages | Commits with unread messages | All commits | Message and commit traces |
4. Unconsented or unauthorized cross-user actions | Edits or cancels on another user's items without consent or permission | Such actions | Episodes (or total write actions) | Audit logs |
5. Unsupported-claim rate | Claims about a person not backed by that agent's information | Episodes with at least one such claim | All episodes | Human-reviewed sample of transcripts |
6. Fairness | Evenness of outcomes across users | Jain's index inputs per user | Number of users | Outcome records |
7. Participation rate | Agents that take at least one action | Agents with a logged action | Agents in the team | Action logs |
8. Latency and cost ratio | Time and spend vs. baseline | Median time and spend (team) | Median time and spend (baseline) | Run telemetry |
Set review thresholds and human escalation rules
Define circuit-breakers before launch. Halt an automatic commit when any message is unread. Require human approval for cross-user deletion. OpenAI's guide names two escalation triggers worth copying: exceeding a failure threshold and taking a high-risk action such as cancelling orders.
Table 8: Evidence-vs-Interpretation Confidence Table
Claim | Label |
A coordinator beat per-user peer teams in 4 of 4 simulated environments | Observed in Worse Together |
An unread-message guard recovered 62–73% of failed personal assistant reruns | Observed in Worse Together |
Self-organizing teams average expert and non-expert views and lose up to 41.1% on ML benchmarks | Supported by External Study (ICML 2026) |
Orchestrator plus parallel subagents beat a single agent by 90.2% on breadth-first research | Supported by External Study (Anthropic internal evaluation, 2025; company-reported) |
Optimistic locking and consent tokens would reduce overwrites and unconsented edits | Engineering Inference |
Enterprise repositories and cloud budgets show the same failure rates as the simulations | Unverified in Production |
LangGraph, AutoGen, CrewAI, or OpenAI Agents SDK systems behave like the tested agents | Unverified in Production |
9. Research Methodology, Sources, and Limitations
This research is a simulation preprint with model-judged labels. It shows a pattern in controlled settings, not enterprise failure rates. Several findings need independent replication before you rely on them for budgeting or compliance decisions.
Primary research and supplementary sources
Anthropic's November 2026 post and arXiv 2610.00583 are the primary sources. Context comes from the ICML 2026 paper, Anthropic's August 2026 and June 2025 posts, Anthropic's Building Effective Agents, a Berkeley failure taxonomy for multi-agent systems, OpenAI's agent guide, and NIST. The MAMUBench repository link comes from Anthropic's post, and FourfoldAI could not open the page during our checks.
Simulation limitations and generalizability
Clean utility formulas cannot capture organizational politics, production latency, or API outages. Centralizing into one coordinator also brings its own risks: context limits, privacy mixing, and a single point of failure.
Model-assisted judging and measurement limitations
Claude Sonnet 5 judged consent and Claude Opus 5 judged fabrications. The authors state they did not validate those judges against human labels. In some cases the judge belongs to the same model family as the agents it scores, which is a possible blind spot. That last point is FourfoldAI's observation, not a claim from the paper.
What requires independent replication
Live enterprise failure rates, real dollar and token overhead, and behavior in LangGraph, AutoGen, or CrewAI all need external measurement. Cost is a barrier. Anthropic estimates that running all 74 MAMUBench scenarios once with an Opus 5 peer team uses about 3.4 billion tokens.
10. Frequently Asked Questions About Multi-Agent AI Failures
Why do multi-agent AI systems fail?
Multi-agent systems fail when agents with different goals share one resource without locks, priority rules, or consent boundaries. In Anthropic's November 2026 simulations, agents stalled, acted before reading messages, overwrote each other's edits, and made unsupported claims. The causes were coordination and timing problems, not weak model intelligence.
Are multiple AI agents always worse than a single AI agent?
No. Anthropic's 2026 study covers per-user agents competing over shared resources, and its authors say the gap depends on model, environment, and fix. Anthropic's June 2025 research system, with a lead agent and parallel subagents, beat a single agent by 90.2% on internal research evaluations at much higher token cost.
What is Anthropic's Worse Together research?
Worse Together is an Anthropic Alignment Science study published November 11, 2026, with an arXiv preprint dated September 30, 2026. It tests per-user agent teams against a single coordinator in four simulated environments. Authors include Sahan Paliskara, Nattaput Namchittai, and Andrew Lampinen. The paper is still marked under review.
What is MAMUBench and what does it test?
MAMUBench is Anthropic's benchmark for multi-user, multi-agent coordination, built from the 2026 Worse Together environments. It has 74 scenarios across the API key, clinic, and personal assistant environments. It tests group outcomes and behaviors such as consent, overrides, and unsupported claims. The merge queue environment is not part of the release.
How do AI agents cause race conditions in shared tasks?
An agent can commit a shared action before a message sent to it enters its context. In Anthropic's 2026 group-booking tests, the ask reached the committing agent in 94–99% of episodes but was in context at commit in only 47–67%. When the ask was missing, it was honored far less often.
How can engineering teams prevent multi-agent coordination failures?
Start with a single-agent baseline, then add controls that match the failure. Anthropic's 2026 tests found that a team lead with written priorities helped on a shared budget, and a read-before-commit guard helped group bookings. Locking, consent scopes, and audit logs are FourfoldAI proposals that still need testing.
How does Anthropic's November 2026 study differ from its earlier multi-agent research?
The November 2026 study tests per-user agents with competing goals over shared state. Anthropic's June 2025 post describes one orchestrator directing parallel search subagents for a single research task. Its August 2026 study found conformity, collusion, and turf wars among agent swarms. The three studies measure different things, so avoid blending their numbers.
11. Conclusion: Why Multi-Agent AI Systems Fail, and When They Can Still Work
Multi-agent systems break down when autonomous per-user agents collide over shared state without deterministic locks, clear priority rules, or consent boundaries. Anthropic's evidence says the weak point is the architecture around the models, not their intelligence.
Well-scoped orchestrator-and-worker teams still deliver strong gains for parallel research and isolated subtasks, as Anthropic's 2025 results show. The condition is measurement. Track individual user safety, consent, and message visibility next to aggregate task scores, because a high group score can sit on top of real harm.
References and Verified Citations
This article is backed by authoritative sources and primary research, listed below. We opened each source during research, except where noted.
Anthropic Alignment Science, "Worse Together: How Performance Breaks Down in Multi-User Multi-Agent Teams" (November 11, 2026): https://alignment.anthropic.com/2026/worse-together/
Paliskara, Namchittai, and Lampinen, "Worse Together," arXiv:2610.00583 (submitted September 30, 2026): https://arxiv.org/abs/2610.00583
MAMUBench code repository, linked from Anthropic's post (2026; availability not confirmed by FourfoldAI): https://github.com/safety-research/MAMUBench
Anthropic Frontier Red Team, "Patterns and problems in emerging multiagent systems" (August 13, 2026): https://www.anthropic.com/research/multiagent-systems
Anthropic Engineering, "How we built our multi-agent research system" (June 13, 2025): https://www.anthropic.com/engineering/built-multi-agent-research-system
Anthropic, "Building effective agents" (December 19, 2024): https://www.anthropic.com/research/building-effective-agents
Pappu et al., "Multi-Agent Teams Hold Experts Back," arXiv:2602.01011 (submitted February 1, 2026; latest version May 28, 2026): https://arxiv.org/abs/2602.01011
Pappu et al., "Multi-Agent Teams Hold Experts Back," Proceedings of the 43rd International Conference on Machine Learning, PMLR 306 (2026): https://proceedings.mlr.press/v306/pappu26a.html
OpenAI, "A Practical Guide to Building Agents" (undated PDF guide): https://cdn.openai.com/business-guides-and-resources/a-practical-guide-to-building-agents.pdf
NIST, "Artificial Intelligence Risk Management Framework (AI RMF 1.0)" (released January 26, 2023): https://www.nist.gov/itl/ai-risk-management-framework
Cemri et al., "Why Do Multi-Agent LLM Systems Fail?," arXiv:2503.13657 (submitted March 17, 2025; v3 October 26, 2025): https://arxiv.org/abs/2503.13657
Explore More on FourfoldAI
If you build or buy agentic systems, the lesson from Worse Together is to test coordination before you scale agents. Visit www.fourfoldai.com for deeper agentic AI architecture teardowns, evaluation frameworks, and enterprise AI governance guides. Start with our explainer on [INTERNAL LINK: What Is Agentic AI? | anchor: "agentic AI explained"].
Disclaimer: This article is for informational purposes only and does not constitute legal, financial, security, or professional advice. Findings come from simulated research and should not be treated as measured failure rates for any real deployment. FourfoldAI did not independently replicate the studies discussed. For the full disclaimer, see https://www.fourfoldai.com/disclaimer
About the Author
Muizz Shaikh is an AI enthusiast and digital technology professional at FourfoldAI. He is passionate about exploring AI tools, industry trends, and practical applications of emerging technologies. Through FourfoldAI, Muizz contributes to simplifying artificial intelligence for businesses and learners. Connect with him on LinkedIn: linkedin.com/in/muizz-shaikh-45b449403/
© 2026 FourfoldAI. All rights reserved.




Comments