AI agents can now use thousands of skills without overwhelming their memory

Researchers built a system that lets AI agents access massive skill libraries (2,000+) without loading everything at once, reducing the tokens needed by 38% while improving task completion by 44%. In practice, this means AI systems handling complex real-world tasks—controlling apps, browsing the web, managing workflows—get faster and cheaper to run.

The constraint has always been simple: an AI agent's working memory fills up fast. Load every possible skill and the system becomes slow, expensive, and hallucination-prone. This paper shows you can solve that by building a dependency map of skills offline, then retrieving only what's actually needed at inference time. The result is that agents can now work with skill libraries an order of magnitude larger than before—the thing that broke the system at 500 skills now works fine at 2,000. That matters because real-world agent tasks are messy and unpredictable; having thousands of specialized tools available, but only loading the relevant few, is how these systems will actually scale beyond toy benchmarks.