Автор: Online Games VIP Editorial
·
IGN reports that four Google AI agents defeated Minecraft’s Ender Dragon after more than 27 hours, coordinating through chat without human intervention.
Four Google AI agents reportedly defeated Minecraft’s Ender Dragon after more than 27 hours of autonomous play. IGN, reporting on a paper from Google researchers, says the agents began in a random Minecraft seed and received only the objective of defeating the game’s final boss.
According to IGN’s report, the agents used the open-source Mineflayer library to control their players and coordinated through in-game chat and shared memory. Researchers could observe the run and stop it if necessary, but IGN says they did not intervene during the attempt.
The result should be understood as a reported research demonstration, not a new Minecraft speedrun record or proof that an AI system has mastered every part of the game. The available report confirms a long autonomous run that ended with the Ender Dragon’s defeat. It does not independently establish every detail beyond the researchers’ account, and several technical questions remain open.
Defeating the Ender Dragon requires considerably more than reaching a prepared arena. IGN explains that the agents had to progress through Minecraft’s normal survival systems, including gathering basic resources, obtaining Ender Pearls and Ender Rods, crafting Eyes of Ender, locating a Stronghold and completing the route into The End.
Google placed four agents into a random seed rather than giving them a pre-scouted world, according to IGN. The researchers reportedly assigned no roles and supplied no prewritten strategy for the run. The stated goal was simply to defeat the Ender Dragon.
IGN reports that the agents developed a division of labor during play. Some built shelters for farming materials from Endermen, while others gathered food or explored. The agents also exchanged combat advice and helped one another recover equipment after deaths. That behavior matters because the experiment was not limited to a single bot following a fixed sequence. It tested whether several agents could coordinate while dealing with a changing game state.
A fifth agent monitored the process for cheating, according to IGN’s account. The paper’s reported setup therefore included four agents playing and another agent overseeing the run. The source does not provide a full technical description of the monitoring system in the supplied material, so the scope of that checking remains something to examine in the paper itself.
IGN reports that the four players exchanged 229 messages, ran 2,783 programs and died 605 times over the playthrough. Those figures show how much trial and error was involved. The agents did not complete the objective through a clean, uninterrupted sequence. They repeatedly lost progress, recovered from failures and continued working toward the same goal.
According to IGN, the group reached The End after about 13.5 hours. Most of the remaining time was spent attempting to defeat the Ender Dragon, with repeated deaths during the boss fight. The report says the full attempt lasted just over 27 hours.
IGN also reports that the agents used 1.35 billion input tokens. The article estimates that this would cost about $1,620 for a normal user, or roughly $60 per hour for the group and $15 per hour for each agent. That estimate is useful context for the experiment, although the supplied report does not provide a separate accounting of the model, infrastructure or pricing assumptions behind the calculation.
The report compares the run with a Minecraft speedrun record of six minutes and 36 seconds. That comparison makes the distinction clear: this was not an efficiency challenge. The agents’ achievement was completing a long sequence without human guidance, rather than completing it quickly.
Analysis: Minecraft gives researchers a visible environment in which a goal can be broken into many dependent tasks. A team cannot normally reach the Ender Dragon without first collecting materials, crafting equipment, navigating the world and surviving enough encounters to keep moving forward. Progress can also be lost through death, which forces players to recover or replace supplies.
That combination makes the game a practical setting for examining behavior over time. An agent may know what the final objective is, but it still needs to remember intermediate plans and respond when circumstances change. A team also needs to share useful information without constantly duplicating the same work.
IGN reports that Google researchers were studying coordination, recovery from failure and sustained pursuit of a goal. The researchers connected those abilities with the broader question of whether AI agents can eventually work reliably alongside people. The experiment therefore uses Minecraft as a controlled setting for studying a more general problem, rather than presenting the game itself as the main product being improved.
The visible nature of the environment is another relevant feature. Players, deaths, inventory changes and movement can be observed in the game world. The researchers could watch the attempt through a live dashboard and pause or stop it, according to IGN. That kind of oversight gives the team a way to inspect what the agents are doing while retaining the ability to end the run.
Still, this result should not be stretched beyond the evidence. A successful Minecraft run demonstrates that the agents completed this reported task under the stated conditions. It does not show that they can transfer the same behavior to an unfamiliar real-world workplace, manage an open-ended human request or operate without technical supervision in every situation.
Analysis: For players, the most interesting part is not that an AI can eventually defeat the Ender Dragon. Human players have been completing that objective for years, and the reported time is far slower than the speedrun comparison cited by IGN. The notable change is that four agents apparently organized their own work, recovered from repeated setbacks and maintained a shared objective for more than a day.
That makes the experiment relevant to the way people already understand Minecraft. The game rewards planning, improvisation and cooperation, but it also creates frequent interruptions. A player may lose equipment, become separated from a group or need to pause one task to gather an overlooked resource. The reported agents faced similar categories of problems, although their internal capabilities and limitations are not fully described in the supplied material.
The result may also encourage players to think differently about what an AI companion would need to do well. A useful companion would have to understand the current objective, communicate a plan, notice when another player needs help and adapt after failure. The reported run offers evidence that a research team is testing those behaviors in a game where mistakes are easy to observe.
It is equally important not to confuse autonomous control with enjoyable teamwork. IGN’s report gives the agents’ message count, program count and deaths, but it does not establish whether their communication was efficient, easy for humans to understand or useful outside the specific experiment. Players would need more evidence before concluding that this kind of system is ready to serve as a reliable Minecraft teammate.
The experiment is built around Minecraft’s survival progression and Ender Dragon encounter, so the game itself is the directly relevant playable reference. You can explore those systems in MINECRAFT.
What remains unconfirmed
The supplied report does not include a complete reproduction guide for the run. It identifies Mineflayer as the open-source library used to control the players and says the agents wrote and debugged their own programs, but it does not list the exact model configuration, prompts, hardware, software versions or full boundaries of the fifth agent’s oversight.
It also does not provide a second-party reproduction of the achievement. IGN presents the result as a claim from Google researchers and reports the paper’s description of the experiment. That is enough to report the development, but the distinction matters: the available evidence is a research team’s account of its own run, not an independently verified tournament result.
The source material also leaves several practical questions unanswered. It does not explain how often the agents received game-state information, how shared memory was structured, whether the agents could inspect their own generated programs in the same way a human developer would, or what specific conditions caused the repeated deaths. Those details would help readers judge how general the result is.
Future evidence worth watching includes the full paper, any released code or logs, and attempts by other researchers to reproduce the run under the same restrictions. It would also be useful to see whether the agents can repeat the result across different seeds, complete the task with fewer failures or cooperate with a human player. Until that evidence is available, the strongest confirmed conclusion is narrower: as reported by IGN, four Google AI agents completed an autonomous Minecraft attempt that ended with the Ender Dragon’s defeat after more than 27 hours.