The setup, described by the developer behind agent-wow, an open-source client built on the AzerothCore private server software, gave the model no rendered frames at all. Instead, the agent built a module to capture dozens of types of server messages, which a Python script polls to build a picture of the world, and sent messages back over the game's network protocol to act. Quest information came from data mining AzerothCore's own SQL files, which the developer likened to a human researching quests on a fan site, except the server's own data is what the server actually runs on.

There was some genuine planning on display, according to the developer: the agent worked through prerequisite quest chains in order, sold junk, equipped upgrades and trained abilities before the zone's final cave, and picked up both cave quests at once. Pathfinding, often the sticking point for game bots, was handled by a C++ helper that plots routes using AzerothCore's navigation mesh files, and the developer called the results optimal, noting the agent even exploited map bugs where collision data was lacking.

The run was on a private server, not the live game, and covered only the first zone. It is also a different approach from the model's earlier Portal run, which used screenshots and took about 24 hours to finish the whole game. The developer's next goals are whether a single agent can reach level 80 unaided, and whether multiple agents can team up through the game's social features, with a longer-term aim of filling a server with AI agents to attempt heroic Icecrown Citadel.