Abstract
Current large language model (LLM) agents rely on a tightly coupled, single-step tool selection mechanism that fails to distinguish between tool capabilities and implementations, reducing flexibility when multiple tools offer similar functions. This disclosure introduces a two-step tool selection framework for agentic LLMs that decouples capability identification from tool implementation. In a first step, the agent uses an LLM to identify the abstract capability needed for a given query. In a second step, a tool selection engine — configured statically via tags or dynamically via LLM prompts — selects the most appropriate tool implementation registered under that capability. This design enables fine-grained tool selection, dynamic adaptability, and operational cost savings. It also introduces clear role separation between ML developers that define agent capabilities and ML administrators that manage tool implementations, improving maintainability and governance. By allowing agents to operate independently of specific tools, the framework enhances precision and flexibility in multi-tool LLM agents.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Jagannath, Kishore, "Tool Selection by Large Language Model (LLM) Agents", Technical Disclosure Commons, (June 17, 2025)
https://www.tdcommons.org/dpubs_series/8240