Abstract

A protocol-level mechanism is proposed herein for multi-round discovery of tools exposed by a tool-providing server to an artificial intelligence (AI) agent without transmitting the agent's task description to the server. A client and server compute independent relevance signals. The client uses task text locally, while the server uses a task-agnostic catalog signal in a first round and an episode signal inferred from shortlisted tool names in later rounds. A union retention rule preserves a tool selected by either the client-local signal or a qualifying server episode signal. Discovery progressively discloses headlines, detail objects, and full definitions using per-round confidence thresholds, with a client-configurable recovery ladder that can reopen candidates, deepen server rankings, escalate to a primary large language model (LLM), and finally fall back to the full catalog. The mechanism reduces transmission and token cost, improves selection from large tool catalogs, and preserves data minimization across organizational boundaries.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Download

Share

COinS