Swarms of OpenAI’s GPT-6.1 Sol agents won 12 of 16 head-to-head games against swarms of Anthropic’s Claude Opus 5.5 agents, by our arithmetic, in experiments Y Combinator’s Labs group ran with Olam Labs. GPT won the Minecraft Bed Wars series 6–3 and the Mindustry series 6–1, according to the YC Labs report published on 29 October.
Y Combinator said on X that it found “civil bureaucracy, thousands of telegraphic A2A messages, and OpenAI models retrieving the game’s source code to beat Claude”. The report describes Olam Labs as a company that “builds multi-agent environments for frontier labs”.
Short explainer · GPT Agent Swarms Won 12 of 16 Games Against Claude
A vertical news Short narrated from this piece’s headline and opening line.
Business Pill · MULTI-AGENT SYSTEMS
A short Business Engineer explainer on multi-agent systems: how several AI agents split and coordinate work. It teaches the general idea only and does not discuss the YC Labs report.
The key insight: In YC Labs and Olam Labs’ report, GPT-6.1 Sol swarms won 12 of 16 games against Claude Opus 5.5 swarms by our arithmetic, and in one Bed Wars game the GPT players sent 1,147 direct messages to Claude’s 82.
Mindustry: GPT in Codex Against Claude in Claude Code
In Mindustry, which the report describes as “a factory-building game with tower defense and real-time strategy”, it says “The GPTs (GPT-6.1 Sol in Codex) won 6-1 against the Claudes (Claude Opus 5.5 in Claude Code) in the 7 games”. It says the series ran for about 9.5 hours.
The report says both teams had “the exact same API to the game, without any special instructions given to them on how to play”. Each team was a pairing of a model and its own coding-agent harness.
Its first finding is headed “GPTs act more like a hive”. The report says the Claudes adopted “a much more democratic โ+1โ system” for proposals, while GPT messages became compressed, with phrases like “G7T5FIRSTcourierLIVE”.
When games got intense and it was ambiguous, the report says GPT agents “would fetch the public code of Mindustry from the web (which was allowed in their sandbox)” to understand units and mechanics, and that “The Claudes never searched up game code”.
Games won by each side in YC Labs and Olam Labs’ head-to-head series: GPT-6.1 Sol swarms won Bed Wars 6-3 and Mindustry 6-1 against Claude Opus 5.5 swarms, per the YC Labs report of October 2026; 12 of 16 is our arithmetic.
What Was Posted on X
In collaboration with @olam_labs, we ran a series of experiments letting agent swarms compete in Minecraft and Mindustry.
We wanted to see what kinds of agent dynamics would emerge when they shared a world over long periods of time.
We found civil bureaucracy, thousands of telegraphic A2A messages, and OpenAI models retrieving the game’s source code to beat Claude.
Y Combinator announced the experiments on X on 29 October 2026, saying it ran them with Olam Labs and found civil bureaucracy, thousands of telegraphic agent-to-agent messages and OpenAI models retrieving the game’s source code to beat Claude. This piece takes its figures from YC’s full report.
“Claudes generally established democracies, GPTs formed hives.”
Daniel Jeffries posted on X on 30 October 2026, putting the line “Claudes generally established democracies, GPTs formed hives” in quotation marks and calling it an awesome read.
Bed Wars: Claude Defended, GPT Went for the Bed
The report says YC ran “nine 45-minute games of Bed Wars” with 10 Claude Opus 5.5 agents against 10 GPT agents.
It says “GPTs won our series 6โ3, usually by breaking Claude’s bed in the final minutes, while Claude played defensively and tried to win on kills”.
In GPT’s three wins in the last two series, the report says Claude was ahead on kills 63–16, 57–19 and 59–24. It says GPT “sent 5-30x more messages”; in game A, GPT players sent 1,147 direct messages and Claude players 82, about 14 times as many, by our arithmetic.
The report adds that Claude “signed its name and said “we” (30% of messages, against GPT’s 5%)”, and that “Claude wrote lessons in private memory files and repeated the mistakes anyway”.
Direct messages in Bed Wars game A, including the lobby: GPT players 1,147, Claude players 82, per the YC Labs report of October 2026; about 14 times as many is our arithmetic.
Fifty Agents and a Ration Ledger
A separate run put “50 GPT-6.1 Sol agents in a survival Minecraft world for 3 hours” with one goal, “Build a thriving civilization”. The report says they “sent about 5,900 messages” and spent the run in crowded shelters, rationing scarce food.
It says “22 of the runโs 34 agent deaths occurred in the first night”, and that “46 of the other 49 agents sent their first message to Ada, the first name on the roster”, who became the hub with 616 direct messages received. “There was no vote, election or challenge”, the report says.
What It Means for Teams Running Agent Swarms
The business lever is model choice for multi-agent work. The report says its setup “tests models the way theyโre used in the real world”, and YC open-sourced the game environments and a fork of the Harbor framework that it says it uses “for a variety of things (enterprises, industrials, etc.)”.
Daniel Jeffries, quoting YC’s post on X, wrote the line “Claudes generally established democracies, GPTs formed hives” in quotation marks. The line is his post’s, not the report’s; the report’s own wording is a “democratic-like decision process” for the Claudes. We covered an earlier essay on swarm behaviour in the Dwarkesh essay on agent civilizations.
The Structural Read
As we read the report, the two families organised differently: it says the Claudes used “a much more democratic โ+1โ system” while GPT messages compressed into shorthand like “G7T5FIRSTcourierLIVE”.
As we read the Bed Wars results, scoring rules shaped the outcome: the report says Claude led on kills 63–16, 57–19 and 59–24 in GPT’s three late wins, and still lost because its bed was broken.
As we read the Mindustry setup, each side was a model inside its own harness, GPT-6.1 Sol in Codex and Claude Opus 5.5 in Claude Code, so the results describe those pairings as deployed.
YC Labs, Agents in Games report, October 2026
“When stuck, GPTs read the source code”
Three Implications
FOR TEAMS RUNNING AGENT SWARMS The report says GPT sent 5-30x more messages in Bed Wars; the report we read gives no token counts or dollar costs.
FOR MODEL SELECTION The report says its setup “tests models the way theyโre used in the real world”, with each Mindustry side a model plus its own coding-agent harness.
FOR EVALUATION YC open-sourced the game environments and its Harbor fork, and lists open questions such as how hierarchies change with team composition.
The analysis puts it in one line: “The surface, not the model, governs autonomy.”
As we read it, the Mindustry series paired each model with its own harness, which is the unit that framework describes. The framework was published in July 2026 and does not discuss this report.
What Is Not Established
The sample is 16 games across two series. In Mindustry each side was a model paired with its own harness, so the results do not separate the model from the harness. The report we read does not give token counts or dollar costs for either side.
The report lists open questions it has not answered, including how hierarchies change with team composition and whether “a large swarm of less capable agents beat a swarm of more capable ones”.
This Week In The AI Supercycle
Execution Fault Lines, Concentrated Demand, and Geopolitical Firewalls
This week’s issue: the AI build is not short of demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.
In YC Labs and Olam Labs’ experiments, GPT-6.1 Sol swarms won 12 of 16 games against Claude Opus 5.5 swarms, by our arithmetic, 6–3 in Bed Wars and 6–1 in Mindustry. The report says the GPT agents messaged far more and read the game’s public source code when stuck, while the Claudes voted with +1 and played defence.
A note on sourcing. Findings come from the YC Labs report “Agents in Games” (October 2026), read in full on 30 October 2026, and Y Combinator’s posts on X; we did not contact YC, Olam Labs, OpenAI or Anthropic. Figures labelled our arithmetic are our own.
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.
