This project studies functional agent calls (FAC), a minimalist framework for building more flexible and expressive LLM agents. Instead of relying on hand-designed scaffolds for search, context management, subtask decomposition, or self-improvement, FAC allows an agent to recreate these patterns dynamically through code and recursive subagent calls. Through this SuperUROP, I will help develop the JAZ implementation of FAC, design and run benchmark experiments across long-horizon and search-based tasks, and analyze how different restrictions on the agent loop affect performance.

I joined SuperUROP because I wanted the chance to spend more time on research that feels genuinely exciting to me. I like that this project sits between building real systems and asking deeper questions about how agents should work. I’m hoping to learn a lot, contribute meaningfully, and help push the project toward something publishable.