OpenAI's GPT-6 Astra completed the orc starting zone in World of Warcraft in 40 minutes with zero deaths — without seeing a single rendered frame. The agent operated entirely on network packets and quest data mined from the server's own files, according to the developer of agent-wow, the open-source client used for the run.

The model played on a private server running AzerothCore, an open-source reimplementation of WoW 3.3.5a — the final build of Wrath of the Lich King. A single prompt in OpenAI's Codex, paired with the agent-wow client, drove the entire session. The accompanying YouTube video describes the run plainly: the agent "starts as a level 1 Orc, completes every quest in the Valley of Trials, and finishes the run in Sen'jin Village."

Agent-wow defines no gameplay mechanics itself — no movement, combat, or interaction logic. Its GitHub repo describes it as an "AzerothCore WoW client designed for autonomous AI agent players" that exposes only a module system for agents to build what they need. The client never touches live Blizzard servers; it runs locally for experimentation.

The developer ran OpenAI's flagship model — released early last month — at extra-high reasoning effort, one setting below maximum. WoW appealed as a testbed because it mixes long-term strategy with short-term tactics. The stated end goal: filling "an entire server with AI agents [to] see if they can clear Icecrown Citadel on heroic difficulty."

Working at the protocol layer

The agent built one module that captures 28 types of server messages, held in memory. A Python script polls those messages to assemble the agent's model of the world and sends messages back to act within it. The developer expected the model would need a higher-level abstraction. "In Practice, it was more than capable of working at the protocol layer," the developer wrote.

For quest knowledge, the agent went straight to AzerothCore's SQL files, pulling quest givers, turn-in targets, and spawn points. The developer likens this to a human player spending hours on Wowhead, the popular WoW database site — with one advantage. The server's own files are the exact data the game runs on, whereas a fan site's records can drift out of date. The project's public workspace instructions also list AzerothCore's source code as an agent resource, though the developer did not confirm whether this run used it.

The agent showed planning behavior beyond simple task execution. It worked through prerequisite quest chains in order, sold junk items, equipped upgrades, and trained new abilities before entering the zone's final cave — and picked up both cave quests at once to complete them together.

Pathfinding from navmeshes

Movement proved the hardest problem, as it does for conventional bots. The agent wrote a C++ helper that plots routes between points using AzerothCore's navigation mesh files, known as mmaps. The Detour pathfinding library computes the route, and the helper returns waypoints as coordinates, or an error when no complete path exists. The developer notes that "from my research into heuristics-based bots, pathfinding is always one of the main challenges," but rates the agent's routing as optimal. The agent even found and exploited map bugs at spots with missing collision properties.

This is not the model's first game. Shortly after release last month, it played through Portal using screenshots plus knowledge of the player's position, completing the entire game over roughly 24 hours. The WoW run used no screenshots at all and covered only the first zone. Other developers have taken different routes into classic games — recent Pokémon Red attempts used a custom small model with a Jev decision-model harness coached by Claude.

The developer's next milestones are a single agent reaching level 80 entirely on its own, and multiple agents teaming up through the game's social features to complete group content — a step toward that heroic Icecrown Citadel goal.