Four Gemini Agents Just Beat Minecraft — But at What Cost?

According to The Decoder, four AI agents running on Google's Gemini 4 Argon recently accomplished a notable milestone by beating Minecraft completely unassisted. Dropped into a randomized world without predefined roles, the agents coordinated strategies, wrote code on the fly, and eventually defeated the Ender Dragon in a continuous run.

What it actually took to win

Autonomy does not mean perfection. Over the course of 27 hours and 14 minutes, the four agents died a combined 605 times. To solve problems dynamically, they wrote 2,783 custom JavaScript programs and exchanged 229 chat messages to warn each other of threats and recover lost items from fallen teammates.

The real-world financial cost

The most striking detail of the experiment is the price tag. Running four frontier agents through complex, long-horizon tasks cost a total of $1,620 at standard pricing rates. While the agents successfully organized themselves without human intervention, solving multi-step business or technical workflows at this scale remains an expensive endeavor.

How multi-agent coordination works

Instead of relying on hardcoded rules, the team deployed identical models and let specialization emerge naturally. One agent focused on resource gathering while others managed defense or tool creation. When mistakes happened, they logged errors in a shared memory pool to avoid repeating them, mimicking human team retrospectives.

Is this ready for corporate workflows?

Not yet for routine tasks. While the experiment proves that frontier models can handle complex, multi-day operations without supervision, the financial cost and high failure rate (over 600 deaths) mean autonomous agents are currently best suited for high-value research or problem-solving where human supervision is too slow or costly.