GPT-6 Astra plays World of Warcraft for the first time with agent-wow
6 hours ago
- LLM agents playing video games like World of Warcraft provide insights into model capabilities without being explicitly trained on the game.
- agent-wow enables AI agents to interact with WoW servers using the network protocol, not computer vision or game hacking.
- A module system allows agents to build custom capabilities, reducing complexity; agents can generate code for missing primitives.
- In a first test with GPT-6 Astra (Codex), the agent created an orc character and completed all starting zone quests in 40 minutes with zero deaths.
- The agent data-mined AzerothCore source code to extract quest requirements and optimal strategies, similar to human research.
- Rather than creating high-level abstractions, the agent worked at the packet protocol layer with a custom gRPC module.
- Pathfinding was handled by a C++ helper using Detour library on AzerothCore's navigation meshes, occasionally exploiting map bugs.
- Future goals include leveling to 80 autonomously, multi-agent coordination, improved observability, and sandboxing for security.