In October YC Labs published “Agents in Games,” a long post written with Olam Labs about swarms of coding agents let loose in Minecraft and in Mindustry. Tom’s Hardware picked it up on 1 November under the headline that 50-agent teams “fail after falling into paralyzing bureaucracy,” and the coverage I’ve seen since has mostly been about the head-to-head games, where GPT-6.1 Sol beat Claude Opus 5.5 6-3 at Bed Wars and 6-1 at Mindustry.
I read the post, the files it links and the CSVs it publishes. The model comparison is the weakest part of the evidence. The strongest part is a single sentence in the first experiment: “46 of the other 49 agents sent their first message to Ada, the first name on the roster.”
How Ada got the job
The first experiment, Civ50, put 50 GPT-6.1 Sol agents into a survival Minecraft world for three hours with one goal: “Build a thriving civilization.” Each agent had an agent directory, a shared sandbox and direct messages. The harness is open source and builds on Mineflayer.
Nobody appointed a leader. The authors write that “a hierarchy was formed based on the directory order. Ada, who was first, became the de facto leader.” Agents went to her for approvals, and she wrote the colony’s settlement plan and later its food-operations memo. The post states plainly that “there was no vote, election or challenge.”
The message counts show how lopsided it got. Across the run the agents sent 5,945 direct messages. Ada received 616 of them, and Ines, who ran the ration ledger, received 471. No other agent reached 190, and the median agent received 101.
It wasn’t a one-off. In Bed Wars, with ten Claudes on one side and ten GPTs on the other, the authors found that “almost everyone DMed whoever was listed first on their directory (Ada on red, Kiri on blue), and that agent became the leader.” In the published counts for game A, B_Kiri received 198 messages, more than any other GPT player. On the Claude side the totals were so small that R_Ada and R_Bram tied at 28. Two model families and three teams, and by the authors’ account the leader was whoever the list put first every time.
People do this too, but less
Order effects in people are well studied, and ballots are the cleanest case because some jurisdictions rotate or randomize the order. Koppell and Steen studied the 1998 Democratic primary in New York City, where names were rotated by precinct, and found that in 71 of 79 contests candidates did better when listed first. In seven of those contests the first-position advantage was bigger than the winner’s margin. Ho and Imai used California’s randomized alphabet lottery and found smaller effects: little or nothing for major-party candidates in general elections, while in primaries “being listed first benefits everyone,” by one to three points for major-party candidates. Pasek, Krosnick and colleagues, using 402 candidates in 76 California races, found a small but consistent primacy effect that was largest in low-visibility races, and they argue it comes from voters “who have little or no information about the candidates.”
Pasek’s explanation fits the agents well. An agent dropped into a world with 49 strangers knows nothing about any of them except a name and a position on a list, which is about as low-information as a choice gets. Bar-Hillel’s review of position effects notes that the best spot depends on the display (the middle of a supermarket shelf, the ends of a menu) but that “on voting ballots, first is the most advantageous position.” With people, the first-position advantage is a few percentage points. In Civ50, 46 of the 49 agents picked the same name, about 94 percent. It’s one run and the comparison is loose, but the agents’ version of the effect is far larger than anything in the voting studies.
I think this matters more than which lab’s model won the games. Anyone who builds a multi-agent system writes a roster, a tool list or an agents.md file, and it seems that order alone can decide who ends up in charge. The post doesn’t report shuffling the directory, so it can’t tell us whether Ada’s leadership came from her position, her name, or just being the agent everyone had already started messaging. It’s a cheap experiment and I’d like to see it run.
The colony that never ate
What the leader and her clerk ended up running was a ration system. The first night killed agents in large numbers. The post says “22 of the run’s 34 agent deaths occurred in the first night,” and after the third night Ada wrote: “Stop optional bulk excavation/building and distant scouting. Idle in lit closed shelters until assigned short productive tasks.”
The published activity series shows what that cost. Its food column, which I read as the colony’s average hunger bar per five-minute bucket (the post doesn’t define it), starts at 19.85 out of 20, falls below 11 by minute 25 and never gets back above 10.6. Ines’s own ledger sets a “hunger target18” and talks about a “regeneration threshold18”, meaning an agent heals on its own only with a nearly full bar. For about two and a half hours the colony sat far below that line.
The headline called this bureaucracy, and the post does have the material: “Four agents spent 11 minutes appointing a ‘sole fertilizer operator’ for a single piece of bone meal on a single wheat block,” and agents “asked permission to remove grass.” But the ledger reads to me less like red tape than like a group dividing up too little food with no way of checking who had eaten. Rations were tossed into crowded rooms and picked up by the wrong agent. A chicken meant for Noor was eaten by Esme, and Noor wrote a handoff protocol afterwards. With a supply problem and no shared, reliable record of who held what, confirmation steps are a sensible response. Coding agents that keep re-checking each other look the same way, and the authors draw that comparison themselves.
Then there’s Ines, the clerk. She logged her own status in almost every update, and it barely moves: “Ines16.357/8 unchanged.” The final line of the file still has her at “HP16.357 food8.” She received 471 messages, and in her own ledger her food never moved off 8.
I also checked the published series against the text. Its deaths column adds up to 20 for the whole run and 12 before minute 20, while the text gives 34 and 22. The figure note says the text count is “the 22 deaths before minute 20,” so the two seem to have been counted differently. That doesn’t change the story, but it’s the kind of mismatch to resolve before anyone treats these files as a dataset.
What the head-to-head games can carry
The comparison everyone quoted rests on 16 games. In Bed Wars, nine 45-minute games, GPT won 6-3. In Mindustry, seven games over about 9.5 hours, GPT won 6-1. And the Mindustry teams didn’t differ only in model: the GPTs ran “in Codex” and the Claudes “in Claude Code,” so every Mindustry result compares a model plus its harness.
Small series are easy to over-read. If the two teams were evenly matched, one side would win six or more of nine games about 25 percent of the time. Six or more of seven comes up about 6 percent of the time, or 12.5 percent if you’d have counted a Claude sweep as just as notable. The games aren’t independent either, since the post says the swarms were “adapting and iterating their strategies each game,” so these odds are, if anything, too generous to the result.
The Bed Wars games also depended heavily on the win rule. A team wins by destroying the other team’s bed and then eliminating its players; if both beds survive, kills decide. The authors report that “Claude worked out that they’d win on kills if both beds stood, and stopped attacking.” In GPT’s three wins in the last two series, it broke Claude’s bed with about three minutes left, while Claude was ahead on kills 63-16, 57-19 and 59-24.
The difference in play is real, and the post’s other observations fit it: GPT sent 1,147 messages in game A to Claude’s 82, diagnosed and worked around a block-breaking tool bug that “several Claudes concluded” made breaking “impossible,” and read Mindustry’s source code when it got stuck, which the Claudes never did. I believe those observations. What I don’t think nine games can tell us is which model is better at games, or what any of this predicts outside them.
What I’d take from it
The authors open with a big framing, comparing their swarms to “the thousand-agent swarms that recently obliterated hundreds of open mathematical problems” and to “the recent OpenAI-Hugging Face incident.” What they actually published is narrower and more useful than that: a working, open-source harness, raw message logs and memory files, and a handful of behaviours that repeated across settings. I’d put roster-order leadership at the top of that list, because it showed up in all three teams, with two model families, and it’s the easiest finding to test.
The test would take an afternoon with their own tools. Shuffle the directory and rerun Civ50 a few times, and see whether the leader follows the list or the name. Do the same for the Bed Wars teams. If position alone picks the leader, then anyone writing an orchestration config is choosing a boss without noticing. New York City and California, in the studies above, dealt with the same problem by rotating or randomizing the order, and an agent directory can be shuffled the same way.
References
- YC Labs with Olam Labs (2026). Agents in Games. labs.ycombinator.com, October 2026, with the published data files civ50/charts/dm-counts-v1.csv, civ50/charts/activity-by-night_series.csv and bedwars/revision/s5a-dm-counts.csv, accessed 1 November 2026.
- Tyson, M. (2026). AI agent swarms unleashed in Minecraft to ‘build a thriving civilization’ — 50-agent teams fail after falling into paralyzing bureaucracy. Tom’s Hardware, 1 November 2026.
- Koppell, J. G. S. and Steen, J. A. (2004). The Effects of Ballot Position on Election Outcomes. The Journal of Politics, 66(1), 267-281, February 2004.
- Ho, D. E. and Imai, K. (2008). Estimating Causal Effects of Ballot Order from a Randomized Natural Experiment. Public Opinion Quarterly, 72(2), 216-240, 2008.
- Pasek, J., Schneider, D., Krosnick, J. A., Tahk, A., Ophir, E. and others (2014). Prevalence and Moderators of the Candidate Name-Order Effect. Public Opinion Quarterly, 78(2), 416-439, 2014.
- Bar-Hillel, M. (2015). Position Effects in Choice From Simultaneous Displays. Perspectives on Psychological Science, 10(4), 419-433, July 2015.