A new framework lets software agents learn procedural skills directly from the multimodal resources humans already create.
Researchers have released RESOURCE2SKILL, a system that ingests tutorial videos, code repositories, articles, and reference artifacts, then distills them into executable skills stored in a hierarchical "Skill Wiki." Each wiki entry combines structured text, code snippets, visual examples, and provenance metadata. At inference time, an agent retrieves relevant skills from the wiki and composes them to complete tasks; if coverage falls short, the system can acquire new skills on the fly. Tested across seven authoring domains, the framework improved average scores by 11.9 percentage points over agents given no skill library, and outperformed strong baseline systems in 26 of 28 benchmark cells.
Most existing agent skill libraries are hand-written or scraped from prior agent runs — which means they inherit whatever gaps or blind spots those sources had. RESOURCE2SKILL taps the enormous existing supply of human-created tutorials and documentation instead, treating the format diversity as a feature: videos capture step-by-step visual operations that plain text misses, while code captures executable patterns that prose descriptions obscure.
The multimodal angle is the genuine novelty here, though "Skill Wiki" sounds like something a product manager named at 4 p.m. on a Friday — the underlying idea, that agents should learn from the same materials humans do, is worth watching as benchmark numbers keep climbing.