Toward a Theory of Agents as Tool-Use Decision-Makers

TL;DR

The paper tackles how to ground autonomous LLM-based agents in a principled epistemic framework. It unifies reasoning and acting as epistemic tools and defines knowledge and decision boundaries, proposing alignment to minimize unnecessary tool use. Key contributions include formal definitions of knowledge and decision boundaries, three guiding principles with lemmas, and a roadmap with training/inference-time alignment and agentic training paths. This framework aims to yield foundation agents capable of adaptive, efficient, and safe goal-directed behavior in open-ended environments.

Abstract

As Large Language Models (LLMs) evolve into increasingly autonomous agents, fundamental questions about their epistemic foundations remain unresolved: What defines an agent? How should it make decisions? And what objectives should guide its behavior? In this position paper, we argue that true autonomy requires agents to be grounded in a coherent epistemic framework that governs what they know, what they need to know, and how to acquire that knowledge efficiently. We propose a unified theory that treats internal reasoning and external actions as equivalent epistemic tools, enabling agents to systematically coordinate introspection and interaction. Building on this framework, we advocate for aligning an agent's tool use decision-making boundary with its knowledge boundary, thereby minimizing unnecessary tool use and maximizing epistemic efficiency. This perspective shifts the design of agents from mere action executors to knowledge-driven intelligence systems, offering a principled path toward building foundation agents capable of adaptive, efficient, and goal-directed behavior.

Figures (6)

• Figure 1: Conceptual framework of agent decision-making based on tool use and knowledge boundaries. The agent is modeled as a goal-directed decision-maker that coordinates internal cognitive tools (e.g., chain-of-thought, reflection) and external physical tools (e.g., actions, models, functions) across a tool use decision boundary. Optimal behavior emerges when tool use decisions align with the knowledge boundary, ensuring the agent invokes only the tools necessary to acquire missing knowledge and efficiently achieve task objectives.
• Figure 2: The tool use decision boundary of agent should align with its knowledge boundary. This alignment represents the optimal behavior of agent that only invoke external physical tools when necessary.
• Figure 3: Training should dynamically adjust the decision boundary relative to the fixed knowledge boundary to optimize efficiency, minimize hallucinations, and prevent tool overuse or underuse.
• Figure 4: Inference-time alignment depends on real-time expansion of the knowledge boundary through interaction, requiring adaptive meta-cognition to balance completeness and efficiency.
• Figure 5: A high-level illustration of Lemma 1.1 for a specific model $m$.