Year: 2026
Type: dissertation
Abstract: LLM-based agents are becoming widespread as tech giants and startups race to integrate AI into their products and break records on benchmarks, but lessons from decades of studying multi-agent systems are not being integrated. Decomposing complex tasks and assigning roles strategically is known to be important for performance of multi-agent systems, including multi-agent LLM systems, but the process of designing these systems is not well studied. In this thesis, we recount established properties of agents, defend that they apply to LLM-based agents, and propose a methodology for optimizing these systems. Specifically, we characterize prompt optimization within the broader machine learning field and then extend existing methods to optimize not just the LLM system prompts but the system structure itself. Initial experiments did not demonstrate performance gains, likely due to methodological limitations and time constraints, suggesting that effectively using role assignment in a difficult, realistic digital assistant environment remains an open challenge. Breaking monolithic agents into smaller, focused components is already exhibiting opportunities for explainability. This new technique opens the way for improving multi-agent LLM systems in an efficient, accessible, and grounded way. More broadly, it is a step in further generalizing and contextualizing the abstractions that have sped up AI development in recent years.
Author: Parent, Zachary Alexander
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7154798198