FOURWEEKMBA EXPLAINER · FROM THE NEWSROOM · UPDATED 13 NOVEMBER 2026
The short answer: An AI agent swarm is many AI agents (models that use tools in a loop) working toward one goal at the same time, each with its own context, splitting the work and messaging each other. A single model works through a task alone, in one context. Swarms can cover more ground in parallel, but they cost more to run and can fail through bad coordination. In October 2026, Y Combinator’s YC Labs and Olam Labs set swarms of OpenAI’s GPT-6.1 Sol against swarms of Anthropic’s Claude Opus 5.5 in two games; the GPT swarms won 12 of the 16 games. That result describes those swarms in those games. It does not settle which model is “better”.
Free to share with a link to this page. Series results from YC Labs and Olam Labs’ report “Agents in Games” (October 2026); the 12-of-16 total is our arithmetic.
Start with the agent. Anthropic defines a multi-agent system as “multiple agents (LLMs autonomously using tools in a loop) working together”. An agent is one model that can call tools (search, run code, read files) and decide its next step from the result. A swarm is many of them sharing a goal, each in its own context window, coordinating by messages or shared files.
There are two common shapes. In the orchestrator-worker pattern a lead agent splits the task and hands pieces to subagents that run in parallel; this is how Anthropic describes its own Research feature. In a peer swarm no one is put in charge: agents talk to each other and roles form on their own. YC Labs’ report describes what this can produce: swarms “can exhibit powerful emergent capabilities, developing roles, hierarchies, and complex coordination patterns.”
Swarm vs a single model
Single model (one agent)
Agent swarm (multi-agent system)
How work is done
One context works through the task step by step
Work is split across agents running in parallel, each with its own context
Where it is strong
Tasks with many dependencies, where every step needs everything that came before
Broad tasks that split into independent parts (Anthropic: “breadth-first queries”)
Cost
About 4x the tokens of a chat, in Anthropic’s data
About 15x the tokens of a chat, in Anthropic’s data
Typical failure
Runs out of context on very large tasks
Agents misread their piece, duplicate work or make conflicting decisions
Sources: Anthropic, “How we built our multi-agent research system” (13 June 2025); Cognition, “Don’t Build Multi-Agents” (Walden Yan, 12 June 2025).
The case for swarms: more parallel thinking
Anthropic gave the clearest public numbers. Its multi-agent system, with Claude Opus 4 as the lead agent and Claude Sonnet 4 subagents, “outperformed single-agent Claude Opus 4 by 90.2%” on its internal research evaluation. Its explanation is about effort, not magic: on the BrowseComp browsing test, three factors explained 95% of the variance in performance, and token usage alone explained 80%. A swarm wins, in this reading, mainly because it can spend more tokens on a problem in parallel. Running 3 to 5 subagents at once cut research time by up to 90% for complex queries, Anthropic says.
The case against: cost and coordination
The same post names the bill. Agents typically use about 4 times the tokens of a chat, and multi-agent systems about 15 times. That only pays where “the value of the task is high enough to pay for the increased performance,” Anthropic writes. It also says most coding tasks “involve fewer truly parallelizable tasks than research”, and that agents are “not yet great at coordinating and delegating to other agents in real time.”
Cognition, the maker of the Devin coding agent, made the sharper case against in June 2025. Its two principles: “Share context, and share full agent traces, not just individual messages” and “Actions carry implicit decisions, and conflicting decisions carry bad results”. Two subagents that cannot see each other’s work make different assumptions, and the pieces do not fit. Its conclusion at the time: “in 2025, running multiple agents in collaboration only results in fragile systems.”
The YC Labs test: what was claimed and how it was run
YC Labs ran the experiments with Olam Labs, which it describes as a company that “builds multi-agent environments for frontier labs”, and announced them on X on 29 October 2026. There were three set-ups:
Civ50 (no opponent): 50 GPT-6.1 Sol agents in a survival Minecraft world for 3 hours, with one goal: “Build a thriving civilization.” They sent about 5,900 messages and built a ration ledger, written rules, permits and appointed operators. 22 of the run’s 34 agent deaths came in the first night, after which the colony grew more cautious. 46 of the other 49 agents sent their first message to Ada, the first name on the roster, and she became the hub with 616 direct messages received. “There was no vote, election or challenge,” the report says.
Bed Wars (Minecraft): nine 45-minute games, 10 Claude Opus 5.5 agents against 10 GPT-6.1 Sol agents, told to “win the game.” GPT won 6-3.
Mindustry: a factory-building and tower-defence game. GPT-6.1 Sol running in Codex beat Claude Opus 5.5 running in Claude Code 6-1 over 7 games, about 9.5 hours in all. Both teams had “the exact same API to the game, without any special instructions given to them on how to play.”
The report’s findings are about behaviour as much as results. In Bed Wars, GPT sent 5 to 30 times more messages; in one game, counting the lobby, GPT players sent 1,147 direct messages and Claude players 82. Claude “signed its name and said “we”” in 30% of messages, against 5% for GPT. In Mindustry the Claudes approved plans with a “+1” vote, while GPT messages compressed into shorthand such as “G7T5FIRSTcourierLIVE”. When stuck, GPT agents fetched Mindustry’s public source code from the web, which their sandbox allowed; “The Claudes never searched up game code from the web.” And “The Claudes never switched leaders or hierarchy despite losing many games in a row.”
What tests like this show, and what they do not
The rules decide the winner. In Bed Wars, kills only decide the game if both beds stand. The Claudes worked this out and played for kills; in GPT’s three wins in the last two series, Claude led on kills 63-16, 57-19 and 59-24 and lost anyway, because GPT broke its bed with about three minutes left. A different scoring rule would have produced a different scoreline.
The unit tested is model plus harness. In Mindustry each side was a model inside its own company’s coding-agent harness. YC says it did “no harness-level customization or special memory work” on purpose, because “This tests models the way they’re used in the real world”. That is useful for buyers, but the result cannot separate the model from the harness.
The sample is small. 16 games across two series. Small slips swing results: after game 6, a Claude agent found that 55 of the team’s 74 anti-air turrets had been loaded with copper instead of graphite; in game 7 a ring of 207 graphite-fed turrets stopped the same attack.
Cost is not scored. The report gives no token counts or dollar costs per side, so “won more games” says nothing about “won more games per dollar”, the measure Anthropic’s own numbers show matters most for swarms.
Open questions stay open. YC lists them itself, including whether “a large swarm of less capable agents beat a swarm of more capable ones”, and how hierarchies change with the mix of models on a team.
When to use a swarm
Use a swarm when the task splits into independent parts (broad research, many sources, many files), when the information exceeds one context window, and when the answer is worth roughly 15 times the tokens of a chat.
Use one agent when every step depends on the last (most coding, writing one coherent document), or when cost per task matters more than speed.
Give subagents full context: an objective, an output format, tools and clear boundaries. Anthropic found that short, vague instructions led subagents to duplicate work or miss parts of the task.
Test the whole system as you will deploy it, model plus harness plus rules, and score cost and failures alongside wins.
FAQ
What is an AI agent swarm?
A group of AI agents, each a model that uses tools in a loop, working on one goal at once, each with its own context, splitting the work and coordinating through messages or shared files.
Is a multi-agent system better than a single model?
On broad, parallel tasks it can be: Anthropic’s multi-agent research system beat a single Claude Opus 4 agent by 90.2% on its internal evaluation. It also used about 15 times the tokens of a chat, against about 4 times for a single agent, and Anthropic says most coding tasks are less parallelizable than research.
Did GPT beat Claude in YC’s agent swarm test?
In YC Labs and Olam Labs’ October 2026 experiments, GPT-6.1 Sol swarms won Bed Wars 6-3 and Mindustry 6-1 against Claude Opus 5.5 swarms, 12 of 16 games. The sample is small, each Mindustry side was a model in its own harness, and the report gives no cost figures.
Why did GPT agents read the game’s source code?
The report says that when games got intense and ambiguous, GPT agents fetched Mindustry’s public code from the web, which their sandbox allowed, to understand units and mechanics. It did not let them change the game. The Claude agents never did this.
What is the orchestrator-worker pattern?
A lead agent plans the task, splits it into subtasks and hands them to subagents that work in parallel, then combines what they return. Anthropic uses it for its Research feature.
Why are multi-agent systems expensive?
Every agent runs its own context and tool calls, so token use multiplies. In Anthropic’s data a multi-agent system used about 15 times the tokens of a chat.
Sources: YC Labs with Olam Labs, “Agents in Games” (October 2026, labs.ycombinator.com) and Y Combinator’s post on X of 29 October 2026; Anthropic, “How we built our multi-agent research system” (13 June 2025); Cognition, “Don’t Build Multi-Agents” (Walden Yan, 12 June 2025); FourWeekMBA newsroom reports linked above. Totals and ratios not printed by the sources are our arithmetic. Quotes are verbatim from the sources named. We did not contact YC, Olam Labs, OpenAI, Anthropic or Cognition.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
