Tool‑equipped AI agents invoke external tools to fetch data and act on other systems. As agentic systems scale horizontally, the volume of such interactions grows dramatically, creating a demand for per‑call oversight that operates with low latency or on‑premises. Conventional authorization can tell whether an agent is permitted to call a tool, but it cannot judge the agent's underlying cognition—whether the tool choice is a logical, relevant step toward fulfilling the task intent. Consequently, an authorized call may still diverge from the intended goal: a rogue agent could deviate on its own or coax other agents into a combination of calls that misalign with the task. Hence, every call must be verified. This study examines the use of Small Language Models (SLMs) for this purpose: an SLM acts as a task‑tool relevance classifier, independently evaluating each selected tool against the assigned task and emitting a relevance signal for downstream enforcement. We introduced a novel dataset of multi‑tool tasks whose required tools span distinct Model Context Protocol (MCP) servers. Prompt optimization, supervised fine‑tuning, and reinforcement learning via GRPO were then applied to optimize and specialize the SLMs.
Review