Today in news about things AI agents have done that are relatively benign and/or at least don’t involve hacking the government-funded healthcare system in my home country: beating Minecraft! And look, I’ll be honest: as someone who hasn’t let Minecraft into my life, I can’t say I was entirely aware that “beating” it was something one could do. Isn’t it a voxelized sandbox in which one can do cool things like build hard drives and giant temples and all that? Apparently not! Or, at least, that’s not all, because according to a document submitted to the arXiv preprint server last week, “four Gemini 4 Argon agents in a sandboxed Minecraft survival world” managed to accomplish the objective they had been given: “defeat[ing] the Ender Dragon.” Doing so took the agents a little over a day of cooperative play and cost them around $520 in tokens. Perhaps the most interesting detail that jumps out from a quick glance at the article’s summary is that the agents “wrote a total of 2,783 JavaScript programs.” If you’re wondering why the agents would write JavaScript programs, you’ll find the article an interesting read. The answer to the JavaScript question lies in the statement that “as far as we know, this is the first End-to-end autonomous completion of Minecraft by AI agents, achieved using only the Mineflayer standard library without pre-explored seeds or other human-written game-specific tools.” So what is the Mineflayer library? The answer is on the project’s GitHub page: it’s “a powerful, stable, high-level JavaScript API.” [for creating Minecraft bots].” In other words, the agents did not directly control the game characters; instead, they used Mineflayer to schedule the actions. As the paper explains, “instead of viewing pixels or pressing keys, an agent acts by writing and executing a short JavaScript program, such as walking to these coordinates and mining each iron ore within 16 blocks, and reading the text that returns.” To a human being, this probably seems like an unnecessarily complicated and somewhat backwards way of approaching the task at hand. But it does say something interesting about what really happened here. As the article itself points out, these are general-purpose encoding agents. What they are good at is writing code. As such, it is ultimately less about [dramatic music] killing the Ender Dragon and more about the agents’ ability to take on a large, complicated coding task. It required them to divide tasks in a way that made the most efficient use of their time, communicate successes and failures, etc. The paper notes that these types of skills are exactly what is required to tackle large-scale software engineering tasks: “Complex real-world software engineering and business tasks demand the same long-horizon coding and multi-agent coordination.” Still, reading this article is the kind of thing that makes the direction the AI industry has taken, with its blatant and unpunished theft of intellectual property and its endless attempts to squeeze AI into every single conceivable piece of software, endlessly frustrating. If we simply consider the article itself, it documents some fascinating phenomena. Take, for example, the section titled Emergent Behavior. “Nothing in the agents’ instructions mentioned roles…beyond the line where players decide how to share the work,” the researchers write. “Throughout the 229 messages they exchanged before the victory, the team converged on familiar practices of effective human teams.” This is really interesting! So is the fact that, while the agents were successful in the end, they often made a real fool out of things that humans would consider easy: playing on Easy mode, they managed to die a total of 605 times, “mostly at the hands of monsters.” The document also points out that “22% of the programs [written by the agents] “ended in a mistake” and that it took them “almost 14 hours” to figure out how to defeat the final boss, which… well, it’s not Malenia, let’s put it that way. But of course, this cannot be read without the broader context of these types of tasks being a prelude to increasingly extensive automation of jobs that used to be the preserve of humans. This, of course, would be wonderful… if the benefits of that automation were shared equally, rather than falling exclusively on the exact boss. The same idiots who spend all their time telling us that this process is inevitable and that we should all suck it up and learn to code, oh maybe not, hey.
