Voyager

Why we included this project

Voyager is a research project that shows what happens when you let a large language model drive an agent that keeps learning on its own. It runs a GPT-4-powered loop inside Minecraft: the agent writes executable code for each behavior it picks up, checks that code against environment feedback and execution errors, and saves successful programs into a growing skill library it can draw on later, even in a fresh world. Because the automatic curriculum, the self-verifying code loop, and the skill store are kept as separate pieces rather than one giant prompt, the system is easier to study and adapt if you want to borrow ideas for your own agent. It also calls the model through a blackbox API instead of fine-tuning, a practical pattern for anyone coupling an agent loop to a hosted LLM. Keep in mind this is research software built around Minecraft, not a drop-in automation tool: you will need to set up a game instance and an OpenAI key before anything runs.