The Google DeepMind Gemini 4 Argon team employed four independent agents to defeat the Ender Dragon in a randomly generated Minecraft survival world without any human intervention. This experiment shows that general-purpose AI models can now demonstrate collaboration and self-correction abilities in more complex gaming environments. The four agents coordinated through in-game chat and shared memory, with distinct roles: janesick crafted tools, gyldenia gathered food, auratian ignited key portals, and bulicame built Enderman shelters and delivered the final blow. The entire process took 27.23 hours, most of which was spent in combat with the Ender Dragon.
Performance of the Four Agents Validates the Claim of "First End-to-End Automation of Minecraft"
The team noted that during the mission, the four agents wrote and executed a total of 2,783 programs, exchanged 229 messages, and died 605 times. Additionally, an agent named The Observer acted as an independent witness to monitor the entire process, checking for cheating and responsible for issuing real-time and summary reports. The entire interaction utilized 135 million input tokens, with a final cost of $810 at Gemini's list price, or $1,620 at standard pricing; these details are outlined in the public research report. Researchers also indicated that this achievement did not require any pre-set seeds or human-written game tools, and was accomplished using the Mineflayer open-source library on a native platform.
This experiment highlights the importance of "consistent planning under long-term conditions" for multi-step tasks and reflects the practical value of multi-agent collaboration and cross-agent knowledge sharing. Previously, strategies for complex games often required specific tools and human design, but today's digital systems can autonomously divide tasks, remind each other, share experiences, and correct pathways. This capability also holds potential value for education, simulation training, and enterprise workflow automation.
At the same time, the research team mentioned that this experiment utilized input data on the scale of 135 million tokens, representing a high-investment end-to-end experiment in terms of cloud resources and computational costs. In terms of pricing and promotion, the tiered strategy for Gemini and the commercialization path for advanced models will influence enterprise-level and individual user demand and investment in advanced reasoning capabilities. From a market observation perspective, Google is maintaining broad access through a tiered approach while enhancing workflow automation and long-term computational performance with advanced models like Argon. This strategy helps improve user retention and the value of cross-product integration.
According to additional information from CNMO, Gemini's tiered strategy reflects not only pricing but also the long-term AI ecosystem layout between the U.S. and China: Argon, as an advanced model, focuses on deeper semantic understanding and long-distance context processing, suitable for business scenarios requiring high precision and stable reasoning; Flash-Lite and Flash continue to serve as options for everyday quick responses and low-cost alternatives, retaining a free tier to attract new users. These changes will help Google further integrate Cloud, Workspace, and video services, enhancing the cross-product user experience and retention rate.
Although this experiment focused on the gaming world, the technological advancements behind it may offer insights for real-world automation workflows and decision-making systems. If Argon can demonstrate significant improvements in more extended and rigorous testing across multiple tasks, the value of AI Pro and AI Ultra will be amplified, and Google will drive deeper integration and collaboration within services like Workspace and YouTube Premium. For long-term development, Google needs to achieve a more stable balance between cost, performance, and user experience.
Currently, users concerned about cost and accessibility can still access stable performance through Flash-Lite and Flash modes, while the advanced reasoning capabilities offered by Argon are aimed at enterprise users requiring longer dialogue histories and complex tasks. The official also emphasizes that users can freely manage or cancel subscriptions and provides guidance on settings and switching; interested readers can visit the relevant pages on ai.google for the latest information. These changes are expected to bring faster responses and resource allocation changes on mobile devices and desktops, potentially impacting daily work and research efficiency.