Agent Voyager

Agent Voyager is an LLM-based agent that learns to play Minecraft by writing, executing, and revising skill code stored in a persistent library — effectively building up a repertoire of procedures it can reuse and combine to solve new tasks.

Unlike systems that learn from fixed datasets, Agent Voyager uses its language model to propose new skills based on failed attempts, then validates whether the new skill actually works before adding it to the library. Each skill is a reusable Python function that the agent can call when facing a similar situation in future gameplay, creating a compounding learning effect where later tasks are solved faster using skills discovered earlier. The agent identifies gaps in its own abilities and designs experiments to fill them, effectively writing its own curriculum.

The quality of the skill library depends on the underlying model's ability to write valid and useful code. Substantial compute is required for continuous skill discovery and validation. Transfer to environments outside Minecraft may be limited, since the specific skills learned are tied to the Minecraft action space and observation format.

How well skills learned in Minecraft generalize to other domains is unclear. Whether the self-curation process eventually plateaus or continues indefinitely is an open empirical question. Whether the approach can be combined with imitation learning from demonstrations to accelerate skill discovery is an active research area.