"Position: Agent Should Invoke External Tools ONLY When Epistemically Necessary" proposes a fundamental shift in how large language models (LLMs) and autonomous agents are designed, trained, and evaluated. Instead of treating tool use simply as a convenient shortcut to maximize task success or reward, the authors argue that agents should only interact with the external world when it is "epistemically necessary"—meaning the agent cannot reliably resolve its uncertainty and complete the task using internal reasoning alone.

To formalize this, the paper introduces the Theory of Agent (ToA), which frames agent behavior as sequential decision-making governed by epistemic constraints. The summary of its core concepts includes:

• Unified View of Reasoning and Acting: Internal reasoning (cognitive tools) and external actions (physical tools) are treated as alternative methods for acquiring knowledge and reducing uncertainty.
• The Knowledge Boundary: Every agent has a latent, model-specific boundary separating the tasks it can solve internally from the tasks that strictly require external interaction.
• Epistemic Effort: Every task has a fixed informational burden that an agent cannot eliminate. The agent can only choose how to allocate this effort between internal cognitive processing and external physical delegation.

The authors highlight that current agent frameworks often suffer from failure modes because they lack principled rules for allocating this effort. Overestimating internal solvability leads to "overthinking" and hallucinations, as the agent relies on internal reasoning when it lacks the required knowledge. Conversely, underestimating internal solvability leads to "overacting" or unnecessary delegation. Crucially, the paper warns that relying on external tools when internal reasoning is sufficient acts as a "reward shortcut"; it not only introduces computational inefficiencies but actively stagnates the development of the agent's internal intelligence by bypassing opportunities for knowledge consolidation.

Ultimately, the paper argues that true agent alignment goes beyond just generating correct answers; it requires effort-consistent decision making. Agents must develop calibrated "meta-cognition" to accurately assess their own capabilities and invoke tools only when justified. To achieve this, the authors suggest new training paradigms, such as next-tool prediction during pretraining and agentic reinforcement learning that penalizes unnecessary delegation.