Dillon DuPont is chief technology officer of Cua, where he develops open-source infrastructure for AI agents that operate computers. His work addresses a practical question behind autonomous software use: whether an agent can interpret an interface, execute the correct action, and prove it completed the task without breaking anything else.
DuPont studied computer science at MIT and worked on projects combining visual computing with physical interfaces. His augmented-reality circuit-design project ARRR.duino received the overall gold award at MIT Reality Hack 2022.
In 2023, he created GPT-4V-Act, a browser agent that translated screenshots and numbered interface elements into mouse-and-keyboard actions. At Microsoft, he contributed to Windows Agent Arena, an open-source desktop-agent benchmark published at ICML 2025. His contributions included webpage and accessibility-tree parsing, benchmark integration, and task evaluation.
DuPont joined Cua in May 2025. He subsequently built tools for recording human computer-use trajectories and developed composite computer-use agents that separate a model’s planning abilities from the visual grounding needed to locate interface elements.
What dependable computer-use agents require
- Reproducible desktop-agent evaluation. Cua-Bench defines each task through its initial computer state, an intended action trajectory, and an evaluator that verifies the outcome. DuPont has authored tasks covering forms, simulated applications, and clipboard-injection detection.
- Professional software exposes capability cliffs. Cua’s collaboration with Snorkel AI evaluates agents using KiCad circuit-design tasks and simulation-based checks. Agents sometimes successfully edited existing schematics but failed to construct circuits from blank canvases.
- Adversarially tested benchmarks. DuPont screens proposed evaluation environments for reward hacking before admitting tasks into a dataset, treating the integrity of the measurement as part of the product.
- Measurable agent world models. Forking and replaying recorded desktop sessions allows developers to compare an agent’s predictions against the computer’s actual state, testing its understanding of the environment independently of whether it eventually completes a task.
