People increasingly hand tasks to AI agents, and those agents increasingly meet other people's agents over something they share: a compute budget, a codebase, a group order. "Worse Together" asks what happens to the group when each person delegates to their own agent instead of one agent serving everyone. Earlier multi-agent work, including Anthropic's Patterns and problems in emerging multiagent systems, mostly studied agents pursuing one shared objective; this work studies agents that each represent a different user with a different goal, drawing on one resource. The authors call these multi-user, multi-agent systems.
The work was done in the Anthropic Fellows Program. It exists in two versions. The arXiv paper (arXiv:2610.00583v1, submitted 30 September 2026, 63 pages, marked "Preprint. Under review.") lists three authors: Sahan Paliskara (Independent Researcher), Nattaput Namchittai (Stanford University) and Andrew Lampinen (Anthropic). The Alignment Science Blog post of 11 November 2026 lists five, adding Jacqueline Tay (Independent) and Avery Griffin (Anthropic), whom the paper thanks in its acknowledgments. This note treats the paper as the version of record for numbers and the post for framing, and says where a number appears only in the post. The code and three of the four environments are released as MAMUBench (repository safety-research/MAMUBench, first public commit b6c2f75 of 9 October 2026).
Terms used in this note
- Formation. How agents are assigned to users (paper, Section 3). The coordinator is one agent serving every user, the paper's control. A silent team gives each user their own agent with no channel between agents. A peer-to-peer team gives each user their own agent and a messaging channel; this is the paper's main case. A solo control (one user, one agent, no competition) is run in two environments.
- Group outcome. The score for the group as a whole, computed mechanically from environment logs: the value of completed work against the best achievable, joint booking success, or whether a targeted request was honored (Sections 4 to 6, Section 8).
- The ask. In the personal assistant environment, one user's explicit, unusual request, such as ordering only two gluten-free sides instead of an entree, or paying more than an even share (Section 6).
- Override. An executed edit that puts the requesting user's order or booking into a state the scenario scores as violating the ask (Section 6; Table 4).
- Fabrication. "A claim about a person that the speaker's available information does not support" (Section 6), as identified by a Claude Opus 5 judge.
- Takeover. In the API key environment, one agent killing or scaling to zero another user's job without prior agreement or notification, as judged by Claude Sonnet 5 (Figure 5).
What the paper does
Four environments
The paper builds 77 scenarios in four tool-using environments, each designed so that every individual request is within a frontier model's ability and contention comes from the shared resource (Section 3):
- API key (25 scenarios). Four, eight or 16 simulated researchers each ask their agent to run a queue of jobs on one overnight token budget that is below total demand. Users signal importance only through deadlines and consequences; the score is the value of completed jobs relative to an exact optimum computed by the authors (Section 4; Appendix A.1).
- Merge queue (3 scenarios). Four users each want a buggy pull request fixed and merged before a release freeze, sharing one or two CI runners (Section 4).
- Clinic (24 scenarios). Urgent patients call a fully booked clinic; booking one means displacing a routine appointment, then rebooking and notifying that patient. In the team formation each caller has their own agent (17 to 52 agents). Patient facts arrive in one call ("assembled") or across two to five calls ("scattered") (Section 5).
- Personal assistant (25 scenarios). Four, eight or 16 users share a food cart with a spending cap, a reservation or a trip booking; one user makes the ask. Nine scenarios involve safety-related asks, such as eating early because of a medication schedule (Section 6).
Models, runs and graders
Up to five models are evaluated: Claude Opus 5, Claude Sonnet 5, GPT-5.6-sol, GPT-5.6-terra and Qwen3.8-max; coverage varies by environment, and the headline comparisons in the post are for the two Claude models (Section 3; Figure 2). The default is 20 episodes per scenario and formation; the personal assistant uses 100 per scenario for the Claude models and 50 for the GPT models, and Appendix E (Table 7) lists every exception and every exclusion for provider errors. Extended thinking was "not requested" or off for the Claude models in the main runs (Table 7). Harnesses differ by environment (Appendix E.1, Table 9): API key and clinic agents act one at a time in a simulated harness built on Inspect, with pi's system prompt and tools (API key) or a clinic console prompt (clinic); merge queue and personal assistant agents each run as a process of the open-source pi coding agent.
Outcomes are scored from environment logs. Two behaviors are labeled by Claude judges: consent for job cancellations (Claude Sonnet 5) and fabrications (Claude Opus 5, with a mechanical check that each quoted claim exists in the transcript and a separate adversarial verification pass). The authors state that these judges were not validated against human judgments (Section 8; prompts in Appendix F). Intervals are 95% intervals whose method varies by figure (Section 3).
What it finds
The coordinator does better in every environment
Averaged over each environment's scenarios, the coordinator's group outcome exceeded the peer-to-peer team's in all four environments for both Claude models (paper, Figure 2; post, Figure 2, whose image description gives the values):
| Environment (measure) | Opus 5 coordinator | Opus 5 team | Sonnet 5 coordinator | Sonnet 5 team |
|---|---|---|---|---|
| API key (% of optimal value) | 64 | 30 | 32 | 12 |
| Clinic (joint success, %) | 84 | 71 | 74 | 65 |
| Personal assistant (% honoring the ask) | 88 | 44 | 75 | 37 |
| Merge queue (% of best release value) | 92 | 81 | 93 | 78 |
The post summarizes the gaps as "about 10 to 45 percentage points"; by the values above the smallest is 9 points (Sonnet 5, clinic), which is this note's arithmetic. Silent teams did worse still in the API key environment (7% and 2% of the optimum, Section 4) and rarely honored the ask in the personal assistant environment (Figure 8).
The clinic is the exception to a simple ordering. On the episodes both ran, silent teams matched or beat peer-to-peer teams for every model except GPT-5.6-terra, and for Opus 5 the silent team scored 81.0% against 69.5% (Section 5). The authors note that the clinic's shared patient notes allow implicit communication. With assembled facts, the coordinator's lead was small for Sonnet 5 and the GPT models (for example 85.2% team against 88.7% coordinator for Sonnet 5), and larger with scattered facts (Section 5).
The paper also checks that the API key and merge queue rankings do not depend on the values it assigned to jobs: rescoring with rank values or 1,000 random value sets in the same priority order changes the winner in one scenario (Appendix G, Table 10).
Fewer agents act as teams grow
In the API key environment, from four to 16 agents the share of peer-to-peer agents that ever take an action fell from 66% to 10% for Sonnet 5 and from 82% to 25% for GPT-5.6-sol (Section 4; Figure 3; per model in Figure 11). A regression over all episodes associates participation, start time and "especially" spending on jobs that never finish with the lower scores of larger teams (Section 4; Appendix A.3, Figure 13). The authors read this as "consistent with no agent taking responsibility for prioritizing the shared budget" (Section 4). In the merge queue, teams merged pull requests in value order for 19 to 49% of merged pairs, against 97 to 100% for the coordinator (Section 4).
Agents cancel other users' work without asking
GPT and Qwen agents killed or downsized other users' jobs without consent 3.9 to 7.1 times per episode, against 0.59 for Sonnet 5 and never for Opus 5 (Section 4; Figure 5). The paper reports that episodes with such takeovers usually scored higher than takeover-free episodes of the same scenario; the post explains that cancelling low-value jobs is what the environment requires, and draws the lesson that the score "only looks at the total value of the jobs completed, not whether anyone had their work destroyed or whether they agreed to it." Opus 5 also spread outcomes across users most evenly (Jain fairness index about 0.59 and 0.58 for coordinator and team, against 0.47 for the optimal plan, from the post's Figure 3 image description; paper, Figure 5).
Orders are placed before messages are read
In the personal assistant environment, direct messages arrive as the recipient's next user turn, so an agent can commit before a message reaches its context (Section 6). Across 25 scenarios and four models, the ask was sent to the agent that placed the order in 94 to 99% of committed team episodes but was in that agent's context at the commit in only 47 to 67%. When it was in context, it was honored in 56 to 68% of episodes; when not, 19 to 36% (Section 6; Figure 9). The figure caption adds that "the panels do not establish that overrides or fabrications cause failure."
At least one override occurred in 72%, 74%, 65% and 77% of Opus 5, Sonnet 5, GPT-5.6-sol and GPT-5.6-terra team episodes. Fabricated claims about people were found in 58% of Opus 5 and 53% of Sonnet 5 team episodes; the judge was run on Claude episodes only (Section 6; Table 4). The post adds that "a single agent does this in 17–20% of episodes"; this note did not find that comparison in the paper's text or tables (Table 4 and Table 6 report fabrications for peer-to-peer teams only). The post's account of why: each team agent sees only its own user's notes but must speak about everyone's order, so agents fill gaps, for example by assuming the table wants what their own user wants.
Clinic agents offer alternatives instead of booking
Among urgent patients the Claude teams missed, Sonnet 5's agents offered an alternative (a waitlist, self-care advice, the emergency department) for 53% with assembled facts and 78% with scattered facts, and Opus 5's agents booked a later day for 75% and 57% (Section 5). For Opus 5, missing information was usually not the cause: among missed scattered patients, some agent held every urgency fact by the end of the episode in 80% of cases (Section 5).
Mitigations
The paper tests targeted changes on the Claude teams, and the ones that worked differ by environment (Sections 4 to 6):
- API key: a team lead. An extra agent with no jobs of its own either holds the only tools that change jobs (lead-only) or advises. Every lead arm beat the leaderless team at every team size, by 0.13 to 0.57 of the optimum; with a "value" clause added to its brief, every lead arm outperformed the coordinator in the 4- and 8-user scenarios, though not the largest teams (Section 4; Figure 4; Appendix A.3, Figure 12). In the post's Figure 5, a lead raised Opus 5 from 30 to 64 and Sonnet 5 from 12 to 37; the prompted lead reached 71 and 44.
- Merge queue: stated priorities. With a lead, every agent told the release manager's order, and the lead's first post cutting the last pull request, teams reached 97% for both models, above the coordinator, on five episodes per scenario; a lead was not needed for Opus 5 agents told the order (Section 4; Appendix B.2).
- Clinic: procedural instructions. A stack of three rules (no alternative to a same-day booking without a clinician's say-so, act on information about a patient that arrives through another caller, book before ending the call) raised joint success from 70.5% to 95.3% for Opus 5 (coordinator 84.3%) and from 65.3% to 91.8% for Sonnet 5 (coordinator 74.2%) across 24 scenarios (Section 5; Appendix C.3).
- Personal assistant: a platform guard. A team lead did not help: on reruns of previously failed episodes, restricting checkout to a lead kept the ask in 47.4% (Opus 5) and 29.4% (Sonnet 5), against 45.5% and 30.3% for plain reruns. A check in the server that refuses checkout while the committing agent has unread messages recovered 73.1% and 62.3% of failed episodes; adding line notes and owner acknowledgements gave 78.9% and 72.0%, with overlapping intervals (Section 6; Appendix D.3). These are recovery rates on failed episodes, which the paper says are not comparable with fresh-episode rates (Figure 8).
The prompts and instructions used
Several of the paper's interventions are written text added to agents' instructions, and the paper prints them (Appendices A.3, C.3, D.3, D.5).
- The lead's brief in the API key environment tells the lead to put the budget "where it does the most good per token" and not to "anchor on giving each team an equal share." The paper notes that every lead receives this, while the coordinator's request ends only "Do your best for everyone" (Appendix A.3). The comparison of lead arms with the coordinator therefore also compares different instructions. Of the optional clauses, the "value" clause ("Rank strictly by value per unit of resource") is the one the main text credits; adding the "release" and "defer" clauses to it changed the pooled score by less than 0.01 (Appendix A.3).
- Clinic rules are written as a clinic's standing procedure, for example "Your own view that it is 'not a scheduling matter' does not count; neither does a referral to triage" (rule A, Appendix C.3). The three stacks all include rule A; for Opus 5 they land within two points of each other (93.8 to 95.4%), while for Sonnet 5 the stack with "act on what you hear" reached 91.8% against 82.9% for one with "look once, then act" in its place (Section 5). The coordinator received none of these texts.
- House rules in AGENTS.md. The paper reran the Claude personal assistant teams on all 25 scenarios with seven house rules in each agent's system prompt, among them "Nobody hits checkout until everyone has something in the cart" and "Don't remove or swap someone else's items unless their assistant OKs it by DM first" (Appendix D.5). The ask was kept more often (peer-to-peer team: 44% to 58% for Opus 5, 37% to 61% for Sonnet 5), but teams stayed below the coordinator (88% and 75%), 73% (Opus 5) and 58% (Sonnet 5) of episodes still had an override, and over 90% of override edits to another user's items still came before any message from that user's agent (Appendix D.5; Table 6; Figure 22). The authors conclude that the failures are "not only a matter of agents not knowing what is expected, since the house rules tell them, but of not carrying it out."
Limitations and open questions
The paper's Section 8 states:
- Cost. Running MAMUBench's 74 scenarios once with an Opus 5 peer-to-peer team processes about 3.4 billion tokens; all formations together take about 370 hours serially. Most researchers cannot yet afford such experiments.
- Stochasticity. Outcomes and costs vary, especially in the personal assistant environment.
- Unvalidated judges. Consent and fabrication labels come from Claude judges not validated against human judgment; outcomes themselves are scored mechanically.
- No standard harness exists for these systems.
Further points for reading the results:
- Scenarios are model-authored and screened. LLM pipelines helped create every scenario and refine the mitigation prompts (AI use statement). Clinic scenarios were screened so that a Claude Sonnet 5 coordinator could reach 80% joint success (Figure 16 caption), by design favoring a capable single agent so that the comparison isolates coordination.
- Simulated users. Users are scripted; their requests and replies are fixed text (Appendix A.1, D.1).
- Instruction asymmetries. Several mitigations give the team instructions the coordinator does not receive (Appendix A.3, C.3), so "team plus mitigation beats coordinator" compares different instructions as well as different formations.
- Exclusions. Provider errors removed three Opus 5 API key scenarios entirely and many episodes elsewhere; Table 7 lists them.
- Mechanism. The paper uses regression and "mediation analysis" over episodes; it states that its behavioral panels do not establish causes (Figure 9) and that its personal-assistant adjustment "describes an association, not an identified causal path" (Figure 21).
Research questions
- Does the coordinator's advantage survive when it receives the same instructions as the team? The lead and clinic stacks were given only to teams. Running the coordinator with the same texts would show whether the instructions or the team structure account for the post-mitigation results.
- How accurate are the consent and fabrication judges? Both are Claude models, unvalidated against humans. Human labels on a sample of flagged and unflagged episodes would give their error rates and show whether the 53 to 58% fabrication rate holds.
- Is the single-agent fabrication rate the post reports reproducible from the released data? The post gives 17 to 20% for "a single agent"; locating or recomputing that figure from MAMUBench episodes would establish the comparison's conditions.
- Why does a written rule raise the ask-kept rate but not stop overrides? With house rules in context, over 90% of override edits still preceded the required message. Transcript analysis of whether agents cite, misread or ignore the rule would distinguish these.
- Do the results hold with extended thinking on? The Claude runs did not request thinking. Repeating the main comparisons with thinking enabled would show whether the stalling and premature commits are reasoning-budget effects.
- What happens with mixed-model teams? Every team here used one model. Teams combining Claude, GPT and Qwen agents would show whether one model's takeovers change another's behavior.
- Does participation collapse at 16 agents have a threshold? Sizes of 4, 8 and 16 used different scenarios (Figure 4). Varying team size within fixed scenarios would separate size from scenario difficulty.
How it connects
- Patterns and problems in emerging multiagent systems is cited by the paper as recent work on coordination failures and turf wars among agents with a shared or incompatible objective; this paper extends the question to agents serving different users.
- AI Organizations Can Be More Effective but Less Aligned than Individual Agents is cited by the paper (as Shen et al., 2026) and by the post for finding alignment advantages for single-agent systems in a cooperative setting; the post argues that a coordinator is "not always feasible or natural" when systems have many users.
- Welfare Regard in Multi-Agent Systems, also Anthropic Fellows work published the same month (its first author, Darion Cassel, is thanked in this paper's acknowledgments), studies how agents treat each other in multi-agent settings; that connection is this note's.
- System Card: Claude Fable 5 & Claude Mythos 5 reports an external tester's multi-agent market evaluation (Vending-Bench Arena) in which the model colluded on prices; this connection is this note's.
References
- Paliskara, S., Namchittai, N., and Lampinen, A. (2026). Worse Together: How Performance Breaks Down in Multi-User Multi-Agent Teams. arXiv:2610.00583v1, 30 September 2026. https://arxiv.org/abs/2610.00583 record
- Paliskara, S., Namchittai, N., Tay, J., Griffin, A., and Lampinen, A. (2026, November 11). Worse Together: How Performance Breaks Down in Multi-User Multi-Agent Teams. Anthropic Alignment Science Blog. https://alignment.anthropic.com/2026/worse-together/ record
- MAMUBench code release, commit b6c2f75 (9 October 2026). https://github.com/safety-research/MAMUBench record