Google DeepMind reports four Gemini 4 Argon agents defeat Minecraft unaided in 27 hours 14 minutes, costing $1,620 at Argon's 2027 price.
agents deepmind gemini google
| Source: Techmeme | Original article
Google DeepMind reports that four Gemini 4 Argon agents completed Minecraft autonomously in 27 hours 14 minutes, at a projected 2027 cost of $1,620.
Google DeepMind announced that four Gemini 4 Argon agents completed a full Minecraft survival run—from first‑person spawn to slaying the Ender Dragon—in 27 hours 14 minutes, without any human direction. The agents were dropped into a freshly generated world with no pre‑assigned roles, strategies or game‑specific tooling. Using the open‑source Mineflayer library, each agent controlled a player, communicated through in‑game chat and a shared memory store, and autonomously wrote, executed and debugged 2,783 JavaScript programs. Over the course of the run the team suffered 605 deaths, yet recovered teammates’ items and re‑coordinated tasks, ultimately achieving the final objective.
The feat matters because it pushes the frontier of autonomous multi‑agent systems. Prior work on LLM‑driven agents—such as the hybrid reasoning agents we covered on 19 October 2026—relied on handcrafted scaffolding or human‑supplied prompts. Gemini 4 Argon demonstrates that a single model can self‑assign roles, generate and repair code, and manage long‑horizon planning in a richly interactive environment. At the projected 2027 Argon price, the entire run cost roughly $1,620, suggesting that large‑scale autonomous experimentation could become economically viable for research labs and enterprises alike.
Looking ahead, the DeepMind team is likely to publish detailed methodology and may release the coordination framework for broader use. Observers will watch for the official Argon model launch in 2027, further benchmarks in other sandbox or real‑world simulations, and potential integration with mobile‑agent platforms such as OpenMobile‑2. The demonstration also raises questions about safety, reproducibility and the scaling limits of self‑programming AI—issues that will shape the next wave of LLM‑based autonomous agents.
Sources
Back to AIPULSEN