Michael JohnsonFour Gemini 4 Argon agents took 27 hours and 14 minutes to defeat the Ender Dragon in Minecraft. In that time, they wrote 2,783 programs and died 605 times. The researchers described the experiment in a paper submitted to arXiv on October 15, 2026.
A significant part of the result was the path to victory: the agents had to maintain a shared goal across many dependent actions, correct mistakes and recover from setbacks. For developers, this is a concrete example of a long-running autonomy test, not simply a demonstration of gaming skill.
How the team was set up
Each of the four agents controlled a separate player through the open-source Mineflayer library. They coordinated through in-game chat and shared memory. According to the authors, people did not assign roles to the agents or write game logic for them: the agents handled planning, program creation and debugging themselves.
The team started in a randomly generated world on easy difficulty. Researchers observed the game through a control panel and could pause or stop the run, but did not need to intervene during execution. The authors report that the agents divided the work, shared information they found and helped retrieve lost items.
What the experiment shows developers
Winning required connecting short-term actions to an overall goal and maintaining coordination after mistakes. The practical interest of the work therefore lies in the observed cycle of “planning — writing code — running — debugging — handing off results”: the team created tools as needed and continued working after numerous deaths.
This cycle is useful to consider when evaluating software agents. Rather than checking for a single successful answer, one can test whether a system maintains its goal across stages, how it recovers from failures, whether it can pass state to other agents and how many attempts it takes to complete a task. These criteria apply to multistep workflows where the result depends on a sequence of operations.
The numbers come from one run under specific conditions: a random world, easy difficulty and control through Mineflayer. The authors describe it as the first autonomous end-to-end Minecraft playthrough by agents in this setup; their work demonstrates that the scenario is feasible and provides a concrete example of long-term coordination. For practical evaluation of agent systems, what matters is the team’s described working process and the cost of recovering from errors, not just the final victory.