Explainable AI Agents: Capture LLM Tool Call Reasoning with Spring AI

When building AI agents with tool-calling capabilities, the LLM's reasoning behind tool selection is typically invisible. Spring AI's Tool Argument Augmenter solves this by dynamically extending each tool's JSON Schema with extra arguments—like an innerThought field and a confidence level—that the LLM must populate when making a tool call. The augmenter intercepts these extra fields before invoking the actual tool, keeping tool implementations clean and unchanged. A practical Patient Health Status Checker example demonstrates how to configure AugmentedToolCallbackProvider, capture reasoning logs, and trace multi-step tool chains where the model self-corrects by calling prerequisite tools first. The captured reasoning can be used for debugging, prompt optimization, audit trails, or feeding into monitoring systems.