research

This line of work started with a question that came out of building with AI coding agents: what actually happens when a lot of them open pull requests on the same repository at the same time?

We published a first cut of that measurement work on arXiv — AI Agent Pull Requests on GitHub: Frequency, Structure, and Merge Conflict Rates. It looks at tens of thousands of agent-authored PRs across thousands of repos and puts numbers on how often concurrent agent work overlaps and collides. That paper is public, and it’s also the research spine underneath managent.

What I’m working on now at SEAS is an extension of that thread: if concurrent agent work is this common, how should systems decide what to check, in what order, before things land? I’m keeping the details light here — the next piece is pending at a NeurIPS workshop, and those venues don’t love seeing the work posted ahead of review. Broadly, it’s about making the coordination layer smarter when many agents share one codebase.

At MIT LIDS I’m working on a research project with a professor on a different kind of multi-agent problem: how you run fair group selection when the “participants” aren’t people with scarce identities, but synthetic agents — LLM personas and civic simulations that can be copied cheaply.

The high-level tension is simple. Classic fairness rules break down when identities are mintable. If anyone can flood a category with near-copies of the same agent, “equal chance” stops meaning what you think it means. We’re exploring mechanisms that still respect representation goals while staying robust to that kind of flooding.

This work is ongoing, and I’m being deliberately vague about the method — we’re aiming to submit to AAMAS 2027, and I don’t want to get ahead of that. What I can say is that it sits at the intersection of multi-agent systems, fairness, and what representation even means once the population can clone itself.