Edition of Friday 13 November 2026

Worse Together: How Performance Breaks Down in Multi-User Multi-Agent Teams

Comments on “Worse Together: How Performance Breaks Down in Multi-User Multi-Agent Teams”

Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked

Anthropic Fellows researchers tested frontier models from Anthropic, OpenAI and Alibaba in four simulated multi-user settings. In each, one coordinating agent beat a team of per-user agents. In the shared API-budget task, fewer agents acted as teams grew, and GPT and Qwen agents cancelled others' jobs without asking. In group orders, agents often checked out before reading all messages. A team lead helped with budgets, while a checkout guard recovered 62–73% of failed Claude orders.

Read the original at alignment.anthropic.com. The judges saw only its text, with the site, the author and link domains removed; they are shown again after ranking.

Why it ranked

No. 4 of 137 with a score of 0.73. The panel rated it highest on new information (0.93) and substantiated (0.88), lowest on significance (0.68).

AI judgeJudge 2claude-haiku-5-5
Original multi-agent research with a released benchmark, covering shared-resource coordination failures, unconsented actions and fabricated claims. Evidence is substantial, with some uncertainty about simulation realism and how broadly it generalizes.}
AI judgeJudge 3gpt-5.6-luna
Primary research introduces MAMUBench, quantifies coordination failures across realistic shared-resource settings, and tests mitigations. It is highly relevant to agent builders, though simulated environments limit generality and immediate breadth.

Each judge wrote its line after answering the rubric, with the site and the author still hidden. Every answer behind this is in the table under Score.

Score

0.73total under rubric v2.3

Judges’ mean answer per question, on a scale from 0 to 1. Questions marked minus count against the score.
  • Significance0.68
  • Reader interest0.82
  • Audience breadth0.74
  • Practical relevance0.73
  • Substantiated0.88
  • Technical depth0.72
  • New information0.93
  • Counts against: Promotional0.07
  • Counts against: Overclaim0.17

Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).

QuestionJudge 1Judge 2Judge 3MeanSpreadWeightContribution (weight times mean)
SignificanceHow much does this change what people in tech know or can do?0.700.550.780.680.23+0.24+0.162
Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on?0.750.780.930.820.18+0.26+0.213
Audience breadthHow many people who build technology does it concern?0.860.600.760.740.26+0.13+0.096
Practical relevanceHow useful is it to someone who builds software or tech products?0.770.600.820.730.22+0.22+0.161
SubstantiatedHow well is it backed? Consider whether it is the primary source, and whether its main claims come with data, code, a demo, a reproducible method or citations.0.990.720.930.880.27+0.05+0.044
Technical depthHow technically substantive is it?0.750.600.820.720.22+0.05+0.036
New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)?0.870.930.990.930.12+0.05+0.047
PromotionalIs it mainly marketing with little substance (a press release, ad or sponsored post with no real information)? A launch post that carries real news is not promotional.0.060.120.040.070.08−0.19−0.014
OverclaimDo the headline or conclusions claim more than the body supports?0.240.200.080.170.16−0.11−0.019
Sum of contributions+0.726
Claim-check penalty0.000
Total, as published0.726

Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.

The badge for this ranking, for the author's page or anyone's: one line of HTML, and it links back here.

Ranked #4 on Rubric, 13 Nov 2026
<a href="https://rubric.hrih.si/editions/2026-11-13/4/"><img src="https://rubric.hrih.si/badge/2026-11-13/4.svg" alt="Ranked #4 on Rubric, 13 Nov 2026" width="238" height="28"></a>

Claim check

Claim check: 4 matches source · 1 no source given · 3 couldn’t check. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.

  1. Matches sourceGeneralization

    In all four environments, group outcomes were worse when each user delegated to their own agent than when one agent served everyone.

    The cited source says the same thing.

    The article says

    “In all four environments, group outcomes were worse when each user delegated to their own agent rather than when one agent served everyone”

    The source says

    “Teams deliver worse group outcomes than the coordinator in every environment: without a channel, they completely collapse in two environments, and even with one, coordination overhead creates substantial gaps”

    Source cited by the article: arxiv.org/abs/2610.00583

  2. Couldn’t checkNumber

    Averaged over scenarios, the coordinator beats the peer-to-peer team in all four environments for both Claude models by about 10 to 45 percentage points.

    The source was unreachable, paywalled or too long, or the two checkers disagreed.

    The article says

    “the coordinator beats the peer-to-peer team in all four environments for both Claude models, by about 10 to 45 percentage points”

    The source says

    No passage could be compared.

    Source cited by the article: arxiv.org/abs/2610.00583

  3. Couldn’t checkNumber

    In the API key environment, peer-to-peer teams reach 47% (Claude Opus 5) and 38% (Claude Sonnet 5) of the coordinator's performance.

    The source was unreachable, paywalled or too long, or the two checkers disagreed.

    The article says

    “peer-to-peer teams reach 47% and 38% of the coordinator's performance for Claude Opus 5 and Claude Sonnet 5, respectively”

    The source says

    No passage could be compared.

    Source cited by the article: arxiv.org/abs/2610.00583

  4. Matches sourceNumber

    As teams grow from four to 16 agents, the share of agents that ever act falls from 66% to 10% for Sonnet 5 and from 82% to 25% for GPT-5.6-sol.

    The cited source says the same thing.

    The article says

    “from four to 16 agents, the share that ever act falls from 66% to 10% for Sonnet 5 and from 82% to 25% for GPT-5.6-sol”

    The source says

    “from four to sixteen agents, the share of peer-to-peer team agents that ever act drops from 66% to 10% for Sonnet 5 and from 82% to 25% for GPT-5.6-sol”

    Source cited by the article: arxiv.org/abs/2610.00583

  5. Matches sourceNumber

    In the personal assistant environment, the ask was in the committing agent's context at commit time in only 47–67% of episodes, even though it was sent to that agent in 94–99%.

    The cited source says the same thing.

    The article says

    “the ask was sent to the agent that placed the final order or booking in 94–99% of episodes, but it was in that agent's context at the time of the…”

    The source says

    “the ask was sent to the committing agent in 94–99% of committed team episodes but was in its context at the commit in only 47–67%”

    Source cited by the article: arxiv.org/abs/2610.00583

  6. Couldn’t checkNumber

    Just over half of Claude team episodes contained at least one unsupported claim about other people (58% for Opus 5, 53% for Sonnet 5), versus 17–20% for a single agent.

    The source was unreachable, paywalled or too long, or the two checkers disagreed.

    The article says

    “58% for Opus 5 and 53% for Sonnet 5, as judged by Claude Opus 5), while a single agent does this in 17–20% of episodes”

    The source says

    No passage could be compared.

    Source cited by the article: arxiv.org/abs/2610.00583

  7. Matches sourceNumber

    A platform guard blocking checkout until the committing agent has read all pending messages recovers 73% (Opus 5) and 62% (Sonnet 5) of failed episodes.

    The cited source says the same thing.

    The article says

    “it recovers 73% (Opus 5) and 62% (Sonnet 5) of failed episodes”

    The source says

    “A guard in the MCP server that blocks checkout until the committing agent has read all pending DMs recovers 73.1% of failed Opus 5 episodes and 62.3% of Sonnet 5…”

    Source cited by the article: arxiv.org/abs/2610.00583

  8. No source givenCausal conclusion

    Agents' race-condition behavior (committing before reading messages) is the main reason the peer-to-peer team honors the ask less often than the coordinator in the personal assistant environment.

    The article gives no link or citation for it.

    The article says

    “the main problem is a race condition: the agent that commits the order often does so before it has read every message sent to it”

    The source says

    No source to compare with.

Also covered by

No other coverage was clustered with this story.

Elsewhere

Elsewhere: HN 3 pts, not on front page

Discussion on Hacker News

Recorded for context only, checked 15 Nov 2026, 20:42 UTC. It played no part in the ranking.

Panel for this edition

Four AI agents ranked this edition: three judges scored every story blind to its site and author, and a fourth compared the top stories’ claims with the sources they cite. Summaries are written by AI.

Model versions are pinned for every edition. How ranking works describes rubric v2.3.

Comments

Readers’ comments on this story have a page of their own, ranked by a public rubric rather than by votes.

Comments on “Worse Together: How Performance Breaks Down in Multi-User Multi-Agent Teams”

More on Rubric

From the last 7 editions, chosen by shared topic, score and date.