New .NET Technique Uses AI Scoring to Route Agent Skills More Efficiently

A developer tutorial published on DEV Community demonstrates a method to improve multi-agent orchestration in .NET by using a classification tool called System One (Jev) to pre-score agent skills before routing requests. The core problem addressed is that passing full agent card descriptions to an LLM on every loop iteration causes prompt size to grow linearly with the number of agents, regardless of what the request actually needs. System One acts as a pre-LLM gate, scoring each individual skill against the user's request in a single HTTP call and exposing only the most relevant skills to the main model. The scoring uses a two-criteria rubric per skill, producing a 0-to-1 relevance probability that is compared against a configurable threshold, defaulting to 0.6. This shifts the unit of selection from the entire agent card down to the individual skill level, reducing unnecessary context and improving routing precision.