45-minute lesson
Persistent capability in fixed-model agents
A systems reconstruction of Voyager’s curriculum, executable memory, and verification-controlled learning loop.
What you will understand
- Curriculum selection
- How current world state and the history of completed and failed tasks shape the next useful objective.
- Executable skill memory
- How verified programs become retrievable, composable behaviors instead of disappearing after one attempt.
- Iterative repair and verification
- How execution errors, environment feedback, and an explicit success check turn code generation into a repair loop.
- Evidence and limits
- What the reported Minecraft results support, what the ablations isolate, and what the benchmark leaves untested.
- Unique items discovered
- 3.3×
- More unique items than the evaluated prior state-of-the-art baselines
- Traversal distance
- 2.3×
- Longer travel distance than the evaluated baselines
- Tech-tree milestone speed
- up to 15.3×
- Fewer prompting iterations to unlock the wooden-tool milestone than AutoGPT
These are reported comparisons from the paper’s evaluated Minecraft setting. The lesson keeps baselines, ablations, citations, and limitations attached to the claims.
Open the Voyager lesson