Overview
- A DeepMind paper posted Oct. 15 reports four Gemini 4 Argon agents autonomously defeated Minecraft’s Ender Dragon by coordinating through the Mineflayer API and writing code for in‑game actions.
- The run took 27 hours and 14 minutes, produced 2,783 generated programs, and recorded 605 agent deaths over the attempt.
- Researchers emphasize key caveats: the world was on Easy difficulty, agents acted via code with access to hidden game state that human players lack, and DeepMind estimated roughly $520 in model usage for the run.
- Separately, a new 35‑billion‑parameter Mixture‑of‑Experts agent called MLExpert raised MLE‑Bench Lite medal rates from 37.9% to 72.7%, showing large gains for autonomous ML engineering within a moderate model size.
- Commercial and tooling moves target cheaper, high‑throughput deployment: Anthropic rolled out Haiku 5.5 (with earlier Sonnet 5.5 and Opus 5.5 updates) for lower‑cost inference and DeepSeek added a web search/fetch API and V4.1‑Flash‑1017 to ground agents in live web data, which could speed production use of agentic systems.