Testing Laya, a Jev-style decision model, on agent skill selection
Using real agent sessions, the author tested the open-weights decision model Laya on picking which skills to load: Jev chose correctly 59% of the time, untrained Laya 18%, and a fine-tuned version 35%.
Experimenting with Laya, the open weights Jev-style decision model on a use case that I think could actually fit these models: agents selecting which skills to load I'm using my real agent sessions and the skills that frontier models picked Jev picked the right skill 59% of the time. Untrained Laya picked it 18% of the time, with a fine-tuned version boosting that to 35% I'm now training the Laya model on ALL of my sessions to see just how high I can get that number! I don't think these models can pick the skills alone, but perhaps they can create a shortlist of 10 skills, then the main model can pick from those instead of 120+