Today in news about stuff that AI agents have done that is relatively benign and/or at least does not involve hacking the government-funded healthcare system in my home country: beating Minecraft! And look, I’ll be honest: as someone who hasn’t allowed Minecraft into my life, I can’t say I was entirely aware that “beating” it was something that one could do. Isn’t it a voxelized sandbox in which one can do cool things like building hard drives and giant temples and whatnot?

Apparently not! Or, at least, that’s not all it is, because as per a paper submitted to the arXiv preprint server last week, “four Gemini 4 Argon agents in a sandboxed Minecraft survival world” succeeded in fulfilling the goal with which they were tasked: “defeat[ing] the Ender Dragon.” Doing so took the agents just over a day’s worth of cooperative gameplay and cost about $520 in tokens.

Perhaps the most interesting detail that jumps out from a cursory glance at the paper’s summary is that the agents “wrote a total of 2,783 JavaScript programs.” If you find yourself wondering why the agents would be writing JavaScript programs at all, you’ll find the paper to be an interesting read. The answer to the JavaScript question lies in the claim that “to our knowledge, this is the first end-to-end autonomous completion of Minecraft by AI agents, achieved using only the standard Mineflayer library without pre-scouted seeds or other game-specific tools written by humans.”

So what’s the Mineflayer library? The answer lies on the project’s GitHub page: it’s a “a powerful, stable, and high level JavaScript API [for creating Minecraft bots].” In other words, the agents didn’t control in-game characters directly; instead, they used Mineflayer to script the actions. As the paper explains, “rather than seeing pixels or pressing keys, an agent acts by writing and running a short JavaScript program, such as walk to these coordinates and mine every iron ore within 16 blocks, and by reading the text that comes back.”

To a human, this probably feels like an unnecessarily complicated and somewhat backwards way of approaching the task at hand. But it does say something interesting about what’s actually happened here. As the paper itself notes, these are general-purpose coding agents. What they’re good at is, well, writing code. As such, this is ultimately less about [dramatic music] slaying the Ender Dragon and more about the agents’ ability to take on a large and complicated coding task.

It required them to divide tasks in a manner that made the most efficient use of their time, communicate successes and failures, etc. The paper notes that these sorts of abilities are exactly what’s required to tackle large-scale software engineering tasks: “Complex real-world software engineering and enterprise tasks demand the same long-horizon coding and multi-agent coordination.”

Anyway, reading this paper is the sort of thing that makes the direction that the AI industry has taken, with its brazen, unpunished IP thefts and its endless attempts to shoehorn AI into any and every conceivable piece of software, so endlessly frustrating. If you simply consider the paper in and of itself, it documents some fascinating phenomena. Take, for example, the section entitled Emergent Behavior. “Nothing in the agents’ instructions mentioned roles… beyond the line that the players decide how to share the work,” the researchers write. “Over the 229 messages they exchanged before the win, the team converged on practices familiar from effective human teams.” This is genuinely interesting!

So too is the fact that while the agents succeeded in the end, they often made a royal bollocks of things that humans would consider straightforward: playing on Easy mode, they nevertheless managed to die a total of 605 times, “mostly to monsters.” The paper also notes that “22% of programs [written by the agents] ended in an error” and that it took them “nearly 14 hours” to figure out how to defeat the final boss, which… well, it’s not Malenia, put it that way.

But, of course, this can’t be read without the more general context of these sorts of tasks being a prelude to ever-more 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, instead of accruing exclusively to the exact same twerps 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. Wait. Maybe not that, hey.