Beyond Output Explanations: Process-Level Legibility for Agentic AI Systems

When LLM-based agents plan, invoke tools, and act over time, users often receive a fluent answer without knowing what process produced it or what consequences it may cascade into. Current explainable AI (XAI) methods tell users what an agent produced, but not how deeply it reasoned, what type of work it emphasized, or why it decided to stop. We call this the process-level explainability gap. This paper makes three contributions: (1) we identify process-level explainability as a missing interface layer for calibrated reliance in agentic AI; (2) we propose three dimensions of process-level legibility—reasoning depth, work-type emphasis, and termination rationale; and (3) we articulate four design principles and propose behavioral drift as a pragmatic failure indicator for process-level understanding. Rather than offering another taxonomy of explanation content, we argue for reorienting the explanation target itself—from outputs to process-stage signals in agent pipelines.