Gemini 4 Agents Beat Minecraft Without Human Help
Four Gemini 4 agents just beat Minecraft with no human intervention, defeating the game’s final Ender Dragon boss entirely on their own. Google researchers revealed the feat in a new paper, showing how general agentic AI models can now complete complex games that previously required custom-built software. The Gemini 4 agents worked individually while sharing tips, threats, and mistakes — a milestone for autonomous AI.
How the Experiment Worked
The team placed four Gemini 4 Argon agents into a sandboxed, randomized Minecraft survival world with a single goal: defeat the Ender Dragon. No roles, strategies, or game-specific tools were assigned. Each agent controlled a player using the open-source Mineflayer library and coordinated with teammates through in-game chat and shared memory.
The setup was deliberately open-ended. Instead of following a script, the agents had to figure out survival, resource gathering, and combat on their own. That open-ended design is what makes the result notable — the agents were not handed a strategy, only a goal.
Agents Split Tasks and Covered for Each Other
During gameplay, the agents quickly elected to split up tasks. They warned each other about threats, gave lessons and tips, provided coordinates, and even recovered items from dead teammates. Each settled into a role:
janesick made tools
gyldenia gathered food and supplies
auratian lit a key portal
bulicame built Enderman shelters and landed the final blow
This division of labor emerged without human direction, with the agents coordinating entirely through in-game chat. The roles were not assigned in advance; the Gemini 4 agents negotiated them on the fly as the situation demanded.
The Numbers Behind the Run
All told, the process took 27.23 hours, with the majority of time spent on the Ender Dragon. According to the team, “The four agents wrote and ran 2,783 programs, exchanged 229 messages and died 605 times along the way.”
A fifth agent called “the Observer” monitored the session as an independent witness — both to check for cheating and to publish live and summary reports. The agents used 1.35 billion input tokens, and the final cost was $810 at Gemini’s list price** or **$1,620 at the standard price.
Those figures underline the scale of the challenge. Hundreds of deaths and thousands of programs show that success came through persistent iteration rather than a single clean plan.
Why This Minecraft Run Matters
The researchers frame the result as a first. “To our knowledge, this is the first end-to-end autonomous completion of Minecraft by AI agents, achieved using only the standard Mineflayer library without pre-scouted seeds or other game-specific tools written by humans,” the team wrote. “Winning required the agents to keep a coherent plan across thousands of dependent steps for more than a day.”
The experiment suggests general agentic AI can now handle long-horizon tasks in open-ended environments — not just narrow, purpose-built challenges like board games. Keeping a coherent plan across thousands of dependent steps is exactly where earlier AI systems struggled. The Gemini 4 agents held that plan together for more than a day.
That has implications well beyond gaming, from multi-agent coordination to sustained planning under uncertainty. Teams of AI agents that can divide work, share warnings, and recover from failure could eventually take on real-world tasks that demand the same skills.
What the Agents Did Right
Several behaviors stand out from the run. The agents:
Communicated constantly — 229 messages kept teammates aligned on goals and dangers.
Recovered from failure — when one agent died, others retrieved its dropped items.
Shared knowledge — lessons and tips passed between agents improved the group’s play.
Adapted roles — no one assigned jobs; the agents chose them based on need.
These are the same coordination skills humans rely on in team games, and seeing them emerge without prompting is the core takeaway.
Four Gemini 4 agents beat Minecraft with no human intervention, splitting roles, recovering lost items, and defeating the Ender Dragon after more than a day of play. It’s a striking demonstration of what general AI agents can accomplish when given a goal and left to coordinate on their own.
