We ran AI agent swarms in Minecraft and Mindustry

labs.ycombinator.com · Oct 29, 2026

YC Labs and Olam Labs ran multi-agent experiments in Minecraft, Bed Wars, and Mindustry to examine how model swarms coordinate, specialize, and compete. The teams used generic agent harnesses, game APIs, Linux sandboxes, scripting tools, shared files, and direct messages; YC Labs open-sourced its Minecraft harness, Multi-Agent Mindustry environments, and a multi-agent Harbor framework fork.

In a three-hour Minecraft survival run, 50 GPT-6.1 Sol agents received the goal of building a thriving civilization. The agents sent roughly 5,900 messages, formed a hierarchy around Ada, the first name in the directory, and created food-rationing, fishing, courier, and permit systems. Twenty-two of 34 deaths occurred on the first night, after which the colony became increasingly risk-averse and focused on survival rather than expansion. YC Labs is also running a separate Minecraft colony of 25 GPT-6.1 Terra agents.

In nine 45-minute Bed Wars matches, 10 GPT-6.1 Sol agents defeated 10 Claude Opus 5.5 agents 6–3. GPT teams sent far more messages, pursued aggressive bed-destruction strategies, and diagnosed a block-breaking tool bug; Claude teams played more defensively and often prioritized kills. In Mindustry, GPT-6.1 Sol agents won six of seven games against Claude Opus 5.5 agents over about 9.5 hours. YC Labs observed denser GPT communication, compressed language, distributed task ownership, and source-code research, while Claude teams more often used stable leadership and proposal voting.

Read original ↗

What people are saying

Commenters were entertained less by the nominal game results than by the agents’ startlingly human-looking organizational failures.

What readers liked

The agents’ improvised rules, role assignment, and post-incident procedures make multi-agent coordination feel both hilariously recognizable and…
Using an LLM to generate deterministic action code can turn slow deliberation into repeatable, lower-error execution rather than forcing every move…

Concerns raised

The agents play Minecraft with a sluggish, almost turn-based rhythm, squandering an open-ended world that should reward quick adaptation and…
Language models appear poorly matched to per-tick game control: each token-level decision adds latency where games and robots need a tight…

View discussion ↗

Read full discussion analysis

Source: Hacker News · Based on 12 sampled comments · Analyzed Oct 30, 2026

Discussion overview

Commenters were entertained less by the nominal game results than by the agents’ startlingly human-looking organizational failures. The Minecraft group’s elaborate procedures for trivial incidents read as funny, unsettling miniature bureaucracy: agents could negotiate roles, document rules, and conduct blame-free reviews, yet spent absurd amounts of time coordinating basic work. Several readers saw that as a vivid demonstration of both emerging social behavior and the distance between language competence and competent real-time action.

The technical discussion converged on latency and control loops as the real bottleneck. LLMs may be capable enough to understand Minecraft’s objectives, but game play and robotics demand continuous perception, simultaneous inputs, rapid reactions, and low-level motor routines. A deliberative model issuing actions or generating scripts behaves more like a turn-based planner than a live player. Some argued that having an agent emit deterministic code is exactly the right division of labor, while others saw dependence on JavaScript harnesses and code-shaped interfaces as evidence that current models are optimized for terminal-like tasks rather than embodied environments. Readers proposed faster models, compact local swarms, and out-of-band behavior scripts as paths toward more responsive agents. There was also enthusiasm for AI-assisted human-versus-human RTS formats and programmable game arenas, though no one suggested these agents are ready for high-tempo competitive play.

Points of disagreement

The sharpest disagreement is whether agents writing deterministic scripts is a strength or an indictment. Supporters argue that an LLM should plan and delegate repeatable execution to code, gaining speed and reliability; critics argue that this workaround exposes a deeper failure to operate fluidly in games and other real-world feedback loops.

What readers liked

The agents’ improvised rules, role assignment, and post-incident procedures make multi-agent coordination feel both hilariously recognizable and surprisingly tangible.
Using an LLM to generate deterministic action code can turn slow deliberation into repeatable, lower-error execution rather than forcing every move through the model.
Programmable RTS and game-server arenas could become compelling human-plus-AI competition spaces, where automation skill is part of the strategy.
Fast-following models, compact local deployments, and shared-GPU swarms have readers imagining affordable persistent NPC communities rather than cloud-only demos.

Concerns raised

The agents play Minecraft with a sluggish, almost turn-based rhythm, squandering an open-ended world that should reward quick adaptation and reactive play.
Language models appear poorly matched to per-tick game control: each token-level decision adds latency where games and robots need a tight continuous feedback loop.
Needing JavaScript harnesses or code-like command interfaces suggests today’s models still navigate terminals better than they inhabit dynamic worlds.
Alphabetical hierarchy and uniform personalities flatten the social experiment, leaving readers wanting richer roles, temperaments, and genuine internal politics.