The experiment was conducted by an AI enthusiast named CozyBlaze, who used the Model Context Protocol (MCP) and a modified SourcePauseTool (SPT) to allow the model to control the game. The game remained paused while the model processed information, and the SPT executed the input commands once the model sent them.
To assist the model, the system provided game screenshots and the player character's location. The model then executed movements and other actions based on its own planned steps. This process explains why the full live stream of the gameplay lasted approximately 24 hours, despite the edited highlight video being only 2 hours long.
The experimenter noted that this achievement should not be viewed as a standardized benchmark for AI capabilities, as many issues still need to be solved. However, he highlighted that a general-purpose agent's ability to autonomously explore and complete a game world aligns with OpenAI's 2016 vision of a single AI agent capable of solving various gaming tasks.
