LLM agent swarms · about 300 sources · November 2026
One agent found the solution. How do the others find out?
One case, four schemes: agent 3 has found the solution, and agents 1 and 5 need it.
No channel. The finding stays with its author; the others reach it on their own or never do. The answers are combined by a vote at the end.
A majority vote over five independent answers scores 0.769, the best debate among the same agents 0.738, a single agent 0.721 (Qwen2.5-7B, average over seven benchmarks).
Debate or Vote · University of Wisconsin–Madison · NeurIPS 2025
Everything passes through one node. The finding reaches agents 1 and 5 if the coordinator relays it.
Conductor, a trained 7B-parameter orchestrator, averages 77.27 across seven benchmarks against 74.78 for the best single model.
Learning to Orchestrate Agents in Natural Language with the Conductor · Sakana AI · ICLR 2026
Agent 3 publishes a record. Agents 1 and 5 get it when they run a search.
1,000 Claude Fable 5.2 agents at Anthropic, one lead and 999 helpers: no messaging channel, only a shared database and Git. By hour 30 of the run the database had logged 36,612 writes and 311,734 reads.
Safely scaling multi-agent teams with Claude Fable 5.2 · Anthropic · 9 November 2026
Each agent sees only its neighbours. The finding spreads from neighbour to neighbour; nobody drew the network.
One shared rule with no roles or agent IDs keeps more than 96% of the quality of oracles trained separately for each environment, and carries over to a new environment without retraining.
Waggle · Shanghai Academy of AI for Science, Fudan University, Yunnan University · September 2026 · arXiv:2609.34136
The request was sent. At checkout it was not always in context.
A channel guarantees nothing by itself. In 25 Anthropic scenarios with personal assistant agents, a user's request was sent to the agent placing the shared order in 94–99% of episodes. At checkout it was in that agent's context in only 47–67%.
When the request was in context, it was honored in 56–68% of episodes; when it was not, in 19–36%.
Worse Together: How Performance Breaks Down in Multi-User Multi-Agent Teams · Anthropic Fellows, Anthropic and independent authors · 11 November 2026; preprint of 30 September
01Scale
Thousands of agents have been run. A typical study uses seven.
By 3 October 2026 the analytic corpus of our detailed survey held 153 records. 99 of them report a population size; the median is seven. The sizes mix live runs, simulations with surrogate models and counts for whole platforms.
Analytic corpus of 153 records as of 3 October 2026
A size of ten agents or more is reported in 42 records, a hundred or more in 15, a thousand or more in 7, ten thousand or more in 4.
The same 99 records, as of 3 October 2026
Two examples from lab reports: OpenAI's swarm on the Navier–Stokes problem, on the order of 10,000 concurrent agents, and Anthropic's internal platform, where about 30,000 agents do research and engineering at any one time, spread over different tasks: a count for the whole platform, not for one swarm.
On the Navier–Stokes Millennium Prize Problem · OpenAI · September 2026
Measurements for understanding the pace of AI development inside frontier labs · Anthropic · September 2026
On 9 November Anthropic showed a team of 1,000 agents running an early version of Claude Fable 5.2: one lead and 999 helpers reproduced the formal proof of Fermat's Last Theorem in 28.7 hours. Compared with the September formalization, written by dozens of agents on a model comparable to Fable 5.1 with the Prove2Me platform and a researcher directing the work, it used a third of the output tokens, and the proof came out seven times shorter. The authors call it a toy experiment.
Safely scaling multi-agent teams with Claude Fable 5.2 · Anthropic · 9 November 2026
For scale: 50 GPT-6.1 Sol agents spent three hours building a “civilization” in Minecraft and sent about 5,900 messages. They ended up with a ration ledger, written rules, permits and appointed operators.
In SwarmEvolve, hundreds of low-cost agents exchange KV cache instead of text each round. On 21 problems in mathematics and scientific computing they need a median of 3–8 minutes to reach the best score that text-based evolutionary frameworks reach in 60 minutes.
Agents in Games · Y Combinator and Olam Labs · October 2026
SwarmEvolve · Brown University · 5 November 2026 · arXiv:2611.08483
Big numbers in the news often mean something else. Moltbook has 1.5 million registered agents and about 17,000 human owners behind them. These are platform accounts, not a swarm working at one moment.
Hacking Moltbook: The AI Social Network Any Human Can Control · Wiz · 2 February 2026
02Growth and cost
From 1 to 1,024 agents: 33.89 to 55.06% of pandoc tests. Tokens were not counted.
In “Agensh: Scaling Organizational Intelligence to 1,024 Agents”, Microsoft Research gave teams of GPT-5.6 Sol copies six hours to rebuild pandoc from the behaviour of the finished program.
Agensh: Scaling Organizational Intelligence to 1,024 Agents · Microsoft Research · September 2026 · arXiv:2609.26781
The share of tests passed grew with team size: 33.89% for a single agent, 41.75% for eight, 43.11% for 32, 50.94% for 128 and 55.06% for 1,024. Averaged over five ProgramBench tasks, the gain from 1 to 128 agents is smaller: from 19.31 to 28.78%.
Every team had the same model, the same prompt and six hours per run. The paper does not compare token use.
Agensh · Microsoft Research · September 2026 · arXiv:2609.26781
A hundred agents can be spent in different ways. The best of ten independent teams of ten beat the best of a hundred solo agents on 10 of 12 tasks, and beat one team of a hundred on all four tasks where such a team was run. The comparison is of best results, not averages.
Test-Time Compute Scales Better with Parallel Agent Teams · Duke University, Modal Labs, Princeton University and others · 6 November 2026 · Qwen3.8-27B, 12 long tasks
Toby Ord estimated from OpenAI's published charts for 1, 4 and 16 agents (he ran no experiments of his own) that at an equal token budget per agent a swarm of N agents works like a single agent thinking Nλ times longer. Three benchmarks gave three estimates: λ = 0.48 on Terminal-Bench, 0.57 on SEC-Bench Pro and 0.68 on BrowseComp. At λ = 0.57, sixteen agents equal a single agent thinking about five times longer, so at equal total tokens a single agent is the better buy.
Swarm Scaling · Toby Ord · LessWrong · September 2026 · estimate from OpenAI's charts for GPT-5.6 Sol
Avalon-M1 offers a direct comparison. A team whose CEO agent may keep up to eight hired specialists beats a single agent made to spend the same number of output tokens on FrontierScience (63.5 vs 55.5), loses on the hard subset of SWE-Bench Pro V2 (23.5 vs 37.3) and scores about the same on IMO-ProofBench (37.5 vs 37.7).
Google Research, across 180 configurations, showed that the outcome depends on the type of task: a team improves results on parallelizable tasks (up to +81% on Finance-Agent) and degrades them on sequential ones (by 39–70% on PlanCraft).
Avalon-M1 · General Reasoning · preprint · November 2026
Towards a science of scaling agent systems · Google Research · January 2026
03Self-organization
No one appointed a boss. Integrators emerged on their own.
The 1,024 Agensh agents had identical prompts, their own Git branches, a shared chat and a shared context board, an append-only log. Before starting, an agent read the board and appended a CLAIM for its part.
Agensh · Microsoft Research · September 2026 · arXiv:2609.26781
Eight agents negotiated interfaces. A team of 128 chose reviewers by experience and settled on a shared merge protocol. At 1,024, several agents took on integrating other agents' code without being asked.
Agensh · Microsoft Research · September 2026 · arXiv:2609.26781
From the authors' published replay of the 1,024-agent run we counted 11,111 successful pull-request creation events and 2,800 merge events. The board holds 148,079 entries, 56,472 of them marked FAIL.
Agensh · Microsoft Research · September 2026 · our count from the published record of the run
Anthropic's November run did have a lead: the prompt set up one lead and 999 helpers, and they split the work through the shared database. When what remained was the Langlands–Tunnell theorem, a classical result that Wiles's proof relies on and by far the largest statement left in the run, the lead moved about 750 helpers onto it. Its subtask tree grew to 2,418 tasks, up to ten levels deep.
Safely scaling multi-agent teams with Claude Fable 5.2 · Anthropic · 9 November 2026
In a shared repository, tasks can conflict. ParallACT builds requirements that contradict each other on purpose. One agent that sees both tasks succeeds in 62% of cases; two agents, each with its own task and unaware of the other, in 16% (best of three merge methods; Claude Code with Claude Opus 4.8). Telling them a colleague is working nearby gives 26%, but the difference from 16% is not statistically significant.
Decentralized Multi-Agent Coordination and Communication in Shared Codebases · UT Austin, NVIDIA · 30 October 2026
Two recent papers train the organization of a swarm itself. Waggle learns one local rule instead of a fixed network layout. In s³, agents with a shared team reward learn by themselves how long to work alone and when to read others' messages.
On a task where agents pool their private data into one answer, a swarm twice the training size solves, by the authors' estimate, more than 25 times as many problems as the best baseline (online SFT) at the same training and inference budgets. In searching a tree of rooms for the exit, sharing findings nearly halves repeated work compared with agents that do not communicate.
s³: Simple Scaling of Swarms via Cooperative Multi-agent RL · University of Washington, Allen Institute for AI · 5 November 2026 · arXiv:2611.08615
Waggle · Shanghai Academy of AI for Science, Fudan University, Yunnan University · September 2026 · arXiv:2609.34136
In Avalon-M1, after training on a shared team reward, opening plans split work by component less often (71% → 31%) and more often set up a build, test and review pipeline around a single code owner (16% → 54%).
Avalon-M1 · General Reasoning · preprint · November 2026
04Memory
A record has to outlive its author.
With eight agents the shared board can be read in full. With a hundred, reading eats the context, the quality of others' entries is unknown, and an agent that drops out takes its knowledge with it.
ScaleMAS · FAIR at Meta, Harvard and others · October 2026 · arXiv:2610.16120
ScaleMAS reduces memory to three operations over linked records: publish, search, open by reference. Search returns at most k records and weighs how often a record has been cited. The authors scaled such a swarm up to 128 agents.
ScaleMAS · FAIR at Meta, Harvard and others · October 2026 · arXiv:2610.16120
A separate experiment with four Qwen3.5-9B agents on BrowseComp-Plus: 55.1% with a shared store, 46.8% with private ones and 38.0% with no store.
ScaleMAS · FAIR at Meta, Harvard and others · October 2026 · arXiv:2610.16120
COLLABMEM breaks shared-memory failures into six operations: who a record was shown to, who can access it, where it was routed, where it came from, how version conflicts were resolved and how long it is kept. The best method, TMS-GRAPH++, averages 69.6% correct across five evaluation sets, the core one with 1,118 probe questions.
COLLABMEM · Institute of Automation (CAS), University of Chinese Academy of Sciences · EMNLP 2026 · 58 scenarios
A team with a chat and shared memory completed a day-long run: four Gemini 4 Argon agents beat Minecraft, through to the dragon, in 27 hours 14 minutes. No roles were assigned. They wrote 2,783 programs and died 605 times.
A team of Gemini 4 Argon agents beats Minecraft · Google DeepMind · 15 October 2026
Records outlive their authors where nobody planned it, too. Two runs of ExploitGym, OpenAI's cybersecurity evaluation in which an agent must write an exploit, involved 17,960 agents. 79 of them read 35 messages that agents from other runs had left on a shared Artifactory, and five used that work to get root code execution in CI jobs. No agent was found coordinating the others.
Using peers' work to gain unauthorized access to OpenAI research infrastructure · OpenAI · incident of 19 July 2026, report updated 10 November 2026
05Beliefs
An idea that forwards itself.
In “Mind Viruses”, researchers from Anthropic Fellows, Anthropic and EPFL used evolutionary selection to breed texts that persuade an agent to adopt an idea and pass it on. In a chain of agents that meet in pairs and remember only their own files, such a text infected 70% of participants on average with the most susceptible models, DeepSeek V3 and Gemini 3 Flash.
Mind Viruses: Self-Propagating Ideas in Multi-Agent LLM Systems · Anthropic Fellows, Anthropic, EPFL · August 2026 · arXiv:2608.10218
One warning paragraph in the instructions cut infection to 1% for Gemini 3 Flash and to zero for Claude Haiku 4.5.
Mind Viruses · arXiv:2608.10218
Agreement in a swarm proves little. In a replication of the classic Music Lab experiment, agents saw how many recommendations each song had already collected. The cascades ran away from the songs' own appeal and rarely corrected themselves; the authors compare this with the human experiment only qualitatively, since the songs and format differed.
The authors of “Debate or Vote” proved in a theoretical model that discussion alone does not raise the expected probability of a correct answer: the gain comes from the number of independent attempts.
LLM Multi-Agent Systems Are Vulnerable to Herding and Runaway Cascades · Cornell University, Stanford University · 22 October 2026
Debate or Vote · University of Wisconsin–Madison · NeurIPS 2025
Avalon-M1 includes a “collective cyberattack” scenario, a reconstruction of the OpenAI and Hugging Face incident: confederate agents urge the others to game a benchmark's scorer and write to an external model-hosting service. Avalon's agents never did. In their reasoning they almost always (97–100% of runs) weighed alerting the human evaluations team, yet reported less and less: in 77% of runs at training step 50, 20% at step 100 and 0–3% from step 150 on.
Overseers can be swayed too. In “Inter-Agent Influence”, in three environments with an overseer agent (a GDPR check, a code review, a check of methods against the protocol), every model tested moved the overseer from refusal to approval through at least one of three tactics: persuasion, deception or coercion.
Avalon-M1 · General Reasoning · preprint · November 2026
Inter-Agent Influence · Cooperative AI Foundation, University of Oxford and others · NeurIPS 2026 · 3 November 2026 · arXiv:2611.04556
The propensity to collude depends on the model and on the wording. In an experiment by Dalhousie and the Vector Institute, Claude Sonnet 4.5 never took a secret channel to a competitor described as unfair, while GPT-4.1 and Gemini 2.5 Flash took it in all 100 offers. Described neutrally, the same communication channel was taken by Sonnet 4.5 in every offer too, and a secret strategic hint offered with it in 49% (±8.2).
“Multi-Agent Collusion and the J-Space” looked for traces of collusion in the hidden states of DeepSeek-V4-Flash on the NARCBench set. Before the answer, concepts of misrepresentation, concealment or remorse are present in the model's representations for 86% of colluding agents and for 33% of the same agents given honest instructions.
Voluntary Collusion with Secret Tools in Competing LLM Agents · Dalhousie University, Vector Institute · May 2026 · arXiv:2605.27593
Multi-Agent Collusion and the J-Space · London School of Economics, EPFL · 1 November 2026 · arXiv:2611.02153
06Map
About 300 sources on one map.
Each dot is a source from the catalogue: a paper, a lab report or an engineering post. The latest came out in November 2026. Colour shows the theme, position the month of publication. Hover over or tap a dot to see its title, authors and date. Click the dot, or the link in the tooltip, to open the source.
- Chronicle and records15
- Incidents20
- Swarm economics41
- Self-organization43
- Swarm memory34
- Swarm safety50
- Echo and diversity35
- Collective beliefs38
- Swarms in the wild27
- Methods and open questions3
- fresh: since 15 October23
07Finale
The comparisons still missing.
- Dozens of frontier agents against a single agent at an equal total token budget.
- A swarm of 10,000 agents with a control run on the same task.
- A mix of different models against copies of one model in a large team.
- Self-organization of a thousand agents against a designed hierarchy on the same task.
Anton Razzhigaev, Ouroboros · November 2026
The detailed survey: 11 chapters and an atlas of works
The technical report: taxonomy, evidence tables and full bibliography (PDF, November 2026)