MEDIATING COMPLEXITY: REASONING LAYERS FROM EMBODIED INTERACTION TO AGENTIC SYSTEMS

Abstract

This dissertation explores how intelligent systems can act as reasoning layers that mediate between human intent and low-level system signals in complex workflows. Across projects spanning embodied interaction, multimodal presentation systems, and machine learning system optimization, I derive empirical findings about how systems can encode interaction through form, couple feedback to user action at the pre-reflective level, fuse heterogeneous signals for intent inference, and expose autonomous reasoning to support human decision-making. These findings are synthesized into DEXAS, a framework for human--agent collaborative reasoning, and instantiated in PROMPTS, an agentic system that turns system traces into diagnostic insights and prunes large configuration search spaces through explainable proposals for optimizing large-scale LLM workloads. Together, this work establishes reasoning layers as a new systems abstraction that reduces the translation burden on experts and reconfigures their role toward high-level judgment.