Most of my research comes back to agents, the harnesses around them, and how much you can get out of a model without making it bigger.
I do all of it in Dr. Sandip Sen's MASTERS research group at the University of Tulsa, which studies multi-agent systems. The five overlap a lot.
What you can get out of agents through orchestration and harness design. This is where my research is heading. I try the ideas out on Synapse, the setup my own agents run on every day.
Jev is TypeSafe's System One model, and it's a really different kind of model. It doesn't write text at all. It answers questions with probabilities, and it's cheap to run. I'm testing where a model like that fits inside a normal agent loop, like filtering the context that comes in or checking the tool calls the agent wants to make. The early results point to fewer tokens used and better protection against prompt injection, and a paper on them is under review. In October 2026 I showed off this research at TU's AIML Graduate Research Showcase, in a talk called Saving the Planet, One Token at a Time. The talk, the code and a Claude Code plugin are on my Jev page.
How LLM agents negotiate and coordinate, including when principled disobedience helps the human. The framework and its domains are on GitHub.
Fine-tuning and orchestration to get more out of small, local models, so we lean on the big ones less. In one unpublished project, a single trained adapter lets a 7B model negotiate anywhere from an aggressive buyer to an aggressive seller, with no retraining.
LODA is an architecture that lets an LLM agent notice when a command conflicts with the goal it was given, and depart from the command when that gets the goal done better. I presented it to researchers at NASA's Jet Propulsion Laboratory during a MASTERS group visit, and we talked about how it could help rovers.
One so far, a workshop paper from AAMAS 2025. The Jev paper is under review.
With Dale Peasley and Sandip Sen. It asks when an agent should override your literal instructions because it understands what you actually want. I presented it at the workshop in Detroit.