What they did
The author built agent-wow, a platform that lets an agent talk to a World of Warcraft 3.3.5a server over the game's network protocol, with no screen reading, no input simulation and no built-in game actions; agents build whatever modules they need through a standard interface. Everything runs on a private AzerothCore server, not Blizzard's. For the first run he gave Codex, using GPT-6 Astra at xhigh, a one-line prompt to create an orc character and complete all quests in the starting zone, then recorded the session from a GM viewpoint.
How it went
Done in about 40 minutes with zero deaths and few complications, then the agent logged out in the next village. Its approach was to read quest requirements, quest givers, turn-in NPCs and spawn coordinates straight out of the AzerothCore SQL files, plan an order, chain prerequisites, sell junk, equip upgrades and train abilities before the cave section, and pick up both cave quests together. It did not build the high-level helpers the author expected; the one module it wrote just sends and subscribes to specific packets, and it worked at the raw protocol level. The full gameplay recording is linked from the post.
Worth knowing
The run was not sandboxed, so the agent could in principle have reached the server database and changed the game; the author draws the line there. HN commenters noted WoW bots are twenty years old and far more efficient; the point here is that a general model did it untrained. The long-term goal is a server full of agents attempting a heroic raid.
GitHub
Discord