Agent Reasoning Tools (ARTs): A Tool Definition Approach for Empower LLM-based Agent Systems

Abstract

The emergence of LLM-based agentic systems is transforming human-technology interaction by enabling proactive collaboration. However, LLMs often rely on external tools due to their lack direct interaction with environments or access to up-to-date information. This makes effective tool definition and management essential, yet such efforts are challenged by rigidity, overhead, and complexity of logic specification. Current methods often focus on external capabilities, overlooking the enhancement of an agent's internal reasoning. This research introduces Agent Reasoning Tools (ARTs) to address these challenges. ARTs are designed to reduce rigidity and overhead while enabling flexible, human-understandable logic definitions that enhance human-AI collaboration. Evaluating ARTs on aspect term extraction shows highly competitive performance. They represent a significant step toward more flexible, transparent, and user-friendly agentic systems.