GPT-6 Astra plays World of Warcraft for the first time with agent-wow. The concept of LLM-powered agents playing video games is fascinating, as they provide insights into how frontier models perform in simulated worlds they were not explicitly trained on. Unlike heuristics-based bots, an LLM agent uses reasoning capabilities to act on the environment like a real player.
I built agent-wow to see how capable frontier models are at playing World of Warcraft. The end goal is to fill an entire server with AI agents and see if they can clear Icecrown Citadel on heroic difficulty. As a starting point, I gave Codex using GPT-6 Astra (xhigh) the prompt: create an orc character and complete all quests in the starting zone. It completed the task easily in 40 minutes with 0 deaths.
World of Warcraft has mechanics that make it an ideal simulated world for evaluating AI, balancing long-term strategy and short-term tactics. Agent-wow does not use computer vision or game hacking; it interacts directly with the game server using the Wo W network protocol. The module system allows agents to build whatever capabilities they need, cutting down complexity by at least an order of magnitude.
Source: Hacker News · Summarized by HeadlinesBriefing