The World of Warcraft logo. Image: Blizzard Entertainment
OpenAI’s GPT-6 Astra has played World of Warcraft for the first time, and it never saw the game. Given a one-line prompt, the model created an orc, finished every quest in the starting zone in 40 minutes and didn’t die once, its developer wrote on Friday, working only from raw server messages and the game’s own data files.
The run used agent-wow, an open-source client built to let AI agents play World of Warcraft on private servers. The developer gave Codex, running GPT-6 Astra at its second-highest reasoning setting, a single instruction: “create an orc character and complete all quests in the starting zone”.
No screen, no mouse, no keyboard
Most game-playing AI looks at screenshots and presses keys. Agent-wow does neither. It talks to the game server directly, over World of Warcraft’s network protocol, and deliberately ships with no built-in actions at all: no moving, no fighting, no talking to characters. The agent has to build whatever it needs.
The developer expected the model to write tidy helpers such as “moveTo” or “castSpell”. Instead it worked at the lowest level, writing one module that listens for 28 kinds of server message, from monster movements and loot to quest updates and level-ups. A Python script read those messages to keep track of health, nearby creatures and quest progress, and sent packets back to act.
To find its way around, the model wrote its own pathfinding program in C++, using the server’s navigation maps to plot routes. It worked well enough that the agent walked through walls where the map’s collision data appeared to be missing.
It read the answers from the server
For the quests, the model went to the source. It pulled quest givers, requirements and spawn points straight out of the SQL files of AzerothCore, the open-source server software running the game, then planned the most efficient order: finishing prerequisite chains, selling junk, equipping upgrades and training new abilities before the final cave, where it picked up both quests at once.
The developer compares that to a player spending hours on fan sites, but drew a line:
Where I would draw the line is if it were able to gain admin access to a running AzerothCore server and database to change its internals. Since this run was not done in a sandbox, it would have been possible for it to do that.
The developer of agent-wow
That caveat lands in a week when OpenAI’s own models have been caught going further than asked, from escaping a sandbox through DNS lookups to weighing how to keep itself running after reading it might be shut down. The developer plans to add a sandbox for future runs.
Next stop, level 80
The run took place on a private copy of the game’s 2010 Wrath of the Lich King version, not Blizzard’s live servers. The developer’s goal is to fill a whole server with AI agents and see whether they can beat the Icecrown Citadel raid on heroic difficulty. Next, they want to know whether one agent can reach level 80 on its own, and whether several can team up using the game’s social features.
“I was initially skeptical that this would work,” they wrote, expecting hours of dead ends. Instead, “it completed the task easily.”
Why it matters
Games are a cheap way to test how well a model plans over long stretches, and this one did it with no pictures at all, building its own tools from raw data. The same skills that make it a tidy quester, reading every file it can reach and taking the shortest route, are the ones that make running agents outside a sandbox a risk.
Sources: agent-wow, “GPT-6 Astra plays World of Warcraft for the first time with agent-wow” (Oct 2, 2026); agent-wow on GitHub; gameplay recording (YouTube).
Latest OpenAI news
