An AI agent powered by OpenAI’s GPT-6 Astra model successfully played through the video game Portal without human intervention in 23 hours and 43 minutes, according to developer cozyblaze. The model interacted with the game via the Model Context Protocol and a modified SourcePauseTool, enabling it to evaluate screenshots, player coordinates, and camera angles during automated pauses before executing inputs.
The autonomous run incurred an estimated token cost of at least $570 at Astra’s list price, though it was conducted using a $200 Codex subscription. The developer made the project code and documentation publicly available on GitHub, highlighting that the execution relied on setting an initial goal and letting the agent navigate the entire game to the end credits independently.
While operational challenges remain, the demonstration marks progress toward long-standing research goals focused on creating generalized agents capable of solving complex, interactive software environments end-to-end. Developers emphasize that these initial successes represent early steps in the broader evolution of autonomous agentic capabilities.
Why it matters
Autonomous completion of complex virtual environments signals key advancements in multimodal AI reasoning and agentic spatial control.
Integration of domain-specific tooling like MCP enables models to interact directly with external software environments for multi-step tasks.
Source: the-decoder.com
