Google researchers say four Gemini 4 Argon AI agents completed Minecraft and defeated the Ender Dragon without any human intervention.
The experiment placed four agents inside a randomized Minecraft survival world and gave them a single goal: defeat the Ender Dragon. Unlike earlier game-playing AI systems that relied on specialized software, the Gemini agents used a general-purpose agentic model.
Agents Created Their Own Roles and Strategy
The researchers did not assign roles, strategies, or game-specific tools to the agents.
Each AI controlled a player through the open-source Mineflayer library and communicated with the others using in-game chat and shared memory.
The agents quickly began dividing work between themselves. They warned each other about threats, shared coordinates, offered tips and lessons, and recovered items from teammates after they died.
Over time, each agent developed a role. Janesick focused on making tools, Gyldenia gathered food and supplies, Auratian activated a key portal, while Bulicame constructed shelters against Endermen and eventually landed the final blow on the Ender Dragon.
The Run Took More Than 27 Hours
The entire experiment lasted 27.23 hours, with most of the time spent working toward the Ender Dragon fight.
During the run, the four agents wrote and executed 2,783 programs, exchanged 229 messages, and died a combined 605 times.
A fifth AI agent, called the Observer, independently monitored the session. Its job was to check for cheating and publish live and summary reports of what happened.
The experiment consumed 1.35 billion input tokens.
Google researchers estimated the cost at $810 using Gemini’s list price, or $1,620 at the standard price.
Researchers Call It a First
The team said this was, to its knowledge, the first end-to-end autonomous completion of Minecraft by AI agents using only the standard Mineflayer library.
The agents did not use pre-scouted world seeds or other Minecraft-specific tools written by humans.
According to the researchers, completing the game required the agents to maintain a coherent plan across thousands of dependent steps for more than a day.
The experiment highlights how general AI agents can now cooperate, divide tasks, recover from mistakes, and complete long, complex objectives without direct human control.
The post Google Gemini 4 Agents Beat Minecraft with No Human Help appeared first on ProPakistani.