PhD Proposal: Mediating Cognition: Intent-Driven Agentic Workflows for Iterative, Multi-Stage Problem-Solving

The increasing complexity of expert domains, such as machine learning performance optimization, highlights a critical limitation of current AI systems: their inability to function as true cognitive partners in iterative, multi-stage problem-solving. Experts are forced to act as "human middleware," constantly translating high-level intent into low-level system actions and managing fragmented information, leading to high cognitive load and suboptimal outcomes. This preliminary proposal introduces the Digital Executive Assistant System (DEXAS), an intent-driven framework designed to overcome this challenge by mediating expert cognition. DEXAS is founded on three interconnected pillars: a Unified Semantic Workspace that integrates heterogeneous data sources (e.g., logs, metrics) into a dynamic, machine-readable state to semantically ground user goals; Collaborative Agentic Workflows that leverage a multi-agent architecture to autonomously decompose problems, delegate tasks, and synthesize solutions under flexible expert oversight; and Longitudinal Context and Reasoning that enables the system to learn from prior interactions and refine its strategies over time. To validate these principles, this dissertation will develop and evaluate the ML Performance Debugging Assistant. This instantiation will investigate three key research questions: (RQ1) the methods for translating ambiguous human intent and diverse data into an effective problem representation for decomposition; (RQ2) the design of multi-agent orchestration for generating and refining solutions with expert-in-the-loop control; and (RQ3) the impact of such an agentic system on expert performance, cognitive load, and trust, including the identification of key interaction design principles for transparency. Through this research, DEXAS seeks to establish a novel paradigm for human-AI collaboration, enhancing expert agency and efficiency in complex digital environments.