posts

Four Gemini agents finish Minecraft by splitting the work

Four Gemini 4 Argon agents, dropped into a fresh Minecraft survival world, killed the Ender Dragon after 27 hours with no human stepping in, according to Google DeepMind. DeepMind and researcher Maxence Faldor say the agents worked in a sandbox through code and text only. Along the way they wrote 2,783 programs, split up jobs over chat, kept shared notes, and debugged each other's errors.

Slaying the dragon is the headline, but the division of labor is what I find interesting. When every action is a program, each attempt becomes something the team can read, reuse, and fix. The shared notes act as memory that outlives one agent's dead end. That is close to the shape of the agent teams people want for real work: break the goal up, hand off, and recover from failures without the need for human intervention. Minecraft is a forgiving lab for this, since a broken script costs a respawn and not a corrupted ledger. If "only code and text" means the game state arrived as structured data rather than pixels, the interface also did much of the perception work a physical or screen-based task would demand.

The post leaves out some specific numbers that would help tell us how far we can scale this specific approach. Was it one run or the best of several? How many of those 2,783 programs failed? How many were unnecessary or duplicates? What was the total token bill for 27 hours of four Argon agents, and how often did agents disagree on work? Those figures separate a coordination result from a long, expensive single success, and I would be interested in seeing the logs to find out which one this is.

Source: x.com/GoogleDeepMind/status/2111127991401247033

← all posts