Google AI Updates: October 19, 2026

1. Four Gemini 4 Argon Agents Beat Minecraft Without Human-Written Game Logic

Google DeepMind. Google DeepMind researchers (Maxence Faldor, Satinder Singh and Roberta Raileanu) put four Gemini 4 Argon agents into a randomly generated Minecraft survival world on Easy difficulty with one goal, defeating the Ender Dragon, and no assigned roles, strategies or game-specific tools. The agents controlled their players by writing JavaScript against the open-source Mineflayer library, coordinated through in-game chat and a shared memory where they logged mistakes, and finished the game in 27 hours and 14 minutes after writing 2,783 programs and dying 605 times. The run cost about $1,620 at the standard Gemini 4 Argon price that takes effect in 2027, and the authors call it the first end-to-end autonomous Minecraft completion by AI agents. For practitioners, it is a concrete long-horizon data point for multi-agent setups where the model writes its own tooling instead of relying on a hand-built harness. Source (paper)

2. Google Research Open-Sources Harness2 for Refining Agent Harnesses Without Labels

Google Research. Harness2 (v0.1.0, github.com/google-research/harness2) improves the prompts, guardrails, memory, tool guidance, subagents, skills, scripts, and plugins around a frozen model during deployment, without verifiers or labeled data. For each task it runs a contrastive Propose, Probe, Compose step that tests a general-purpose and a domain-specific harness edit and keeps only changes backed by observed behavior, then stores lessons in an EXPERIENCE.md file for future tasks. One refinement step improved all four model and harness setups on all three benchmarks in the paper, from +11.5 points (GLM-5 under OpenCode on WorkBuddy-Bench, 64.2 to 75.7) down to +1.2 (GLM-5 under Codex on JobBench). Source