AI Coding Live News

YC Labs pits AI agent swarms in Minecraft and Mindustry: GPT swarms win 12 of 16 head-to-head games against Claude, while a 50-agent colony slides into bureaucracy

YC Labs, a research project at Y Combinator, and multi-agent environment startup Olam Labs published "Agents in Games" on Oct 29. Fifty GPT-6.1 Sol agents told to "build a thriving civilization" in survival Minecraft spent three hours huddled in shelters rationing food, sending about 5,900 messages and inventing ration ledgers, permits and job titles. In nine Bed Wars games, 10 GPT-6.1 Sol agents beat 10 Claude Opus 5.5 agents 6–3; in seven 50-vs-50 Mindustry matches lasting about 9.5 hours, GPT won 6–1. The leader was almost always whichever agent was listed first in the directory, and stuck GPT agents fetched the game's public source code on their own. The Minecraft harness is open source. These are one lab's experimental observations, not a standardized benchmark.

YC Labs pits AI agent swarms in Minecraft and Mindustry: GPT swarms win 12 of 16 head-to-head games against Claude, while a 50-agent colony slides into bureaucracy

Key Takeaways

  • Head-to-head: GPT-6.1 Sol won Bed Wars 6–3 over nine games and Mindustry 6–1 over seven, 12 of 16 overall (see the YC Labs report)
  • Civ50: 50 GPT agents sent about 5,900 messages in three hours; 22 of 34 deaths came on the first night, after which the colony turned cautious
  • No election: 46 of the other 49 agents first messaged Ada, first on the roster, who received 616 DMs and became leader
  • Communication: in one Bed Wars game GPT players sent 1,147 DMs versus Claude's 82; GPT drifted into compressed shorthand while Claude voted with "+1"
  • Open source: the long-running Minecraft agent harness is public at yc-software/minecraft-agent-harness
Read full story

OverSeer: Multi-Agent World Models Beyond What Agents See via Joint Auxiliary Streaming

Researchers introduced OverSeer in arXiv:2610.28554, pioneering a new class of multi-agent world models operating beyond the field of view of individual agents. Standard video world models collapse in multi-agent environments because individual agents operate under partial observability; naive frameworks feeding external global map inputs suffer severe spatial inconsistency and state hallucination. OverSeer resolves this by jointly predicting auxiliary spatial representation streams alongside the agents' first-person camera streams. In comprehensive Minecraft benchmarks, OverSeer achieves decisive state-of-the-art results across physical grounding, long-horizon spatial memory, and cross-view consistency.

OverSeer: Multi-Agent World Models Beyond What Agents See via Joint Auxiliary Streaming

Key Takeaways

  • OverSeer introduced in arXiv:2610.28554, establishing the first scalable multi-agent video world model framework
  • Replaces external static map inputs with auxiliary spatial streams jointly predicted alongside agents' egocentric perspectives
  • Achieves state-of-the-art performance across spatial grounding, long-horizon memory, and multi-view consistency in Minecraft
  • Provides an ungrounded simulation backbone for multi-robot swarm collaboration, multiplayer game AI, and embodied multi-agent fleets
Read full story
Linked Models & Agents:

Google DeepMind: Four Gemini 4 Argon Agents Autonomously Beat Minecraft from Scratch in 27 Hours by Writing 2,783 Programs

Google DeepMind announced that a team of four Gemini 4 Argon agents completed Minecraft from scratch in 27 hours and 14 minutes without human assistance. Operating in a sandboxed survival world using only standard Mineflayer APIs, code execution, and text chat, the agents autonomously divided roles, recovered teammates' items across 605 deaths, and wrote and debugged 2,783 JavaScript programs to defeat the Ender Dragon.

Google DeepMind: Four Gemini 4 Argon Agents Autonomously Beat Minecraft from Scratch in 27 Hours by Writing 2,783 Programs

Key Takeaways

  • First-ever end-to-end autonomous Minecraft completion: four Gemini 4 Argon agents defeated the Ender Dragon from a random survival seed in 27 hours and 14 minutes
  • 2,783 self-written programs with zero human help: coordinated via in-game chat and shared notes, debugging code and recovering items across 605 deaths
  • WANDR record & Perplexity launch: Gemini 4 Argon simultaneously set a new record on the WANDR benchmark and rolled out to Perplexity Pro and Max users
Read full story

IGMWorld: Benchmarking Coding Agents on Executable World Editing across Minecraft and Terraria via Industry-Grade Modding

Interactive world models are progressing beyond passive simulation toward active world generation and agency. However, deliberately editing an existing, running executable world while preserving unedited systemic invariants remains an underexplored frontier. An international research consortium introduces IGMWorld and IGMBench (arXiv:2610.02331), framing world editing through industry-grade game modding across Minecraft (Java) and Terraria (C#/.NET). Spanning 110 tasks evaluated against 1,100+ deterministic executable criteria across property, entity, dynamics, and system-level intervention depths, IGMBench rigorously assesses coding agent precision. Frontier agents achieve 78.2% task-level success and 94.8% criterion-level pass rates. Crucially, failure rates scale monotonically with intervention depth—most failed edits build and load without syntax errors but diverge in runtime execution logic. Project assets are accessible at vinesmsuic.github.io/IGMWorld.

Key Takeaways

  • Introduces IGMWorld and IGMBench, the first benchmark for executable world editing using Minecraft and Terraria game modding
  • Defines four hierarchical intervention depths (property, entity, dynamics, system) with 110 tasks and 1,100+ deterministic executable criteria
  • Frontier coding agents achieve 78.2% task-level and 94.8% criterion-level accuracy, identifying sharp drops in deep system interventions
Read full story
ADSponsored

GPT-6 Astra Sets a New Milestone for AI Agents in Minecraft

LLM evaluator Vals AI documented GPT-6 Astra completing an unusually long Minecraft task: building a semi-automatic blaze farm, collecting blaze rods, and reaching a warped forest for ender pearls. A creeper then destroyed its chest and bed, exposing practical weaknesses in critical-item protection and recovery.

Key Takeaways

  • GPT-6 Astra built a semi-automatic blaze farm and collected six blaze rods
  • It reached a warped forest, defeated multiple endermen, and collected three ender pearls
  • The lost inventory highlights remaining gaps in persistent memory, risk management, and recovery planning
Read full story
Linked Models & Agents: