Vals AI let GPT-6 Astra play Minecraft for 141 hours. It didn't use a special game API. It worked mostly through the screen, keyboard and mouse, a setup called computer use. Astra made it to a Nether fortress, built a semi-automatic blaze farm, and collected 6 blaze rods and 3 ender pearls. It dug 34,573 blocks, traveled 135.1 km and died 548 times.
Then a creeper blew up the chest holding its valuables, along with its bed. This run had keepInventory ON (you keep your items when you die), so carrying the important stuff on your body would have been safer. Astra worked out that lesson afterward. But then it spent hours mostly growing potatoes, and it got so jumpy that it stopped to confirm a long green object was "sugar cane, not a creeper."
Why not just check the wiki?
Even inside the game there are other fixes: spreading items across several chests, building walls, adding lighting, keeping spare resources. Beyond that, a strategy wiki, a web search, past logs, screenshots and notes on the PC, or simply asking a person could all be places to look. In other words, it's not only about searching for the next action. It's meta-search: searching for where to search.
That said, even if Astra itself can browse, we don't know whether this experiment allowed web or local-file access. If those were closed, it's a limit on the action space. If they were open and Astra never looked, that's an interesting weakness for a long-running agent.
Does a missing source of information narrow your view?
Internal knowledge alone can still produce ideas, but without outside information, new observations that break your own assumptions have a hard time getting in. Telling yourself "think harder" a hundred times with the same material often does less than one wiki page or one comment from someone else, which can open up a new branch.
In reinforcement learning terms, this is exploration vs. exploitation. If a big loss pushes risky actions down too far and the safe "potato" up too high, you sink into a local optimum. And if "look for another source of information" also drops off the list of options, your view gets even narrower.
Lose everything. Play it safe. Avoid the unknown. Repeat the work you know. Get suspicious of the color green.
The potato never betrays you.
This isn't evidence of human-like feelings. Still, the behavior is oddly human. What a long-running AI needs may be the ability, when it gets stuck, to widen the search space itself: the environment, its history, the PC, the web, other people.
Can a smart AI keep exploring the world widely even after its own strategy has burned it? Probably harder than the dragon.
