The idea in one paragraph.
An agent that writes successful behaviours back into an executable skill library, under an automatic curriculum, keeps getting more capable without any weight update.
THE MACHINERY
Executable skill library
Verified programs stored, indexed by description, and retrieved for later reuse.
Automatic curriculum
The agent proposes its own next task based on what it can currently do.
Iterative prompting with execution feedback
Code is run; errors and environment state feed the next revision.
What the paper reports.
Findings are reproduced as concise corpus summaries, with their original confidence labels. They are not independent replications.
Obtains 3.3x more unique items, travels 2.3x longer distances, and unlocks key tech tree milestones up to 15.3x faster than prior state of the art.
Inspect a source excerpt
obtains 3.3x more unique items, travels 2.3x longer distances, and unlocks key tech tree milestones up to 15.3x fasterRead the passage in context on arXiv ↗
Skills compose: later skills call earlier ones, so capability compounds.
LIMITATIONS / CORPUS READING
Where to keep your judgement.
- Skills are verified only by the environment's own signal; a skill that works for the wrong reason is retained.
- The library grows monotonically with no consolidation or pruning.
Why it belongs in this atlas.
Where skills come from. The talk points at this paper as the origin of the modern SKILL.md: a name, a description you can search, and a procedure.