From Black Box to Toolbox: Human-Centered Explainability in Developer-Authored Agent Rules
Abstract
The promise of tailored agent behavior is undermined by a critical explainability challenge: it is difficult to assess how closely and consistently the agent follows user-defined rules. As Large Language Models (LLMs) transition from static assistants to autonomous agents, developers have pioneered markdown-based rule files (e.g., GEMINI.md, CIDER_AGENT.md) to steer agent behavior and mitigate a "organizational context gap" that emerges when general-purpose models lack the "organizational context" necessary for contextually relevant results. This paper presents a qualitative study of 12 Google software developers (n=12) to investigate the authoring and efficacy of these agent rules. Our findings reveal that while rules are intended as technical steering mechanisms, they function as a "Black Box" of validation, where 12/12 participants rely on anecdotal "vibe checks" due to a profound lack of formal evaluation and explainability frameworks. We identify this opacity as a systemic Attribution Gap, which prevents developers from discerning whether a successful outcome was the result of deliberate logic or "pure luck." Paradoxically, these files serve a dual role as "Living Documentation," bridging technical instruction for AI with sociotechnical onboarding for humans. We argue for a transition toward library-level governance and rigorous traceability to transform agent customization from an ad-hoc craft into a human-centered science by revealing the internal "seams" of rule interpretation.
Research Areas
