Graham Neubig

Graham Neubig is a Carnegie Mellon University researcher in natural language processing and large language models, and a cofounder of All Hands AI, the company building OpenHands. His contributions include collaborative work on the DyNet neural-network toolkit, the WebArena benchmark, and OpenHands, alongside a public argument for building capable software-development agents in the open.[3][6][7][12][11][10]

Neubig joined OpenDevin early, after Binyuan Hui and Junyang Lin created the project. In his March 2025 retrospective, he dated its beginning to March 12, 2024 and described how he, Robert Brennan, and Xingyao Wang became early contributors, subsequently formed All Hands AI, and renamed the project OpenHands.[7]

His early contribution was a front end that gave people something to try while the agent was still taking shape. In an October 2024 founder interview, Brennan recalled finding the repository when it contained only a README, then seeing Neubig commit an interface resembling Devin. Neubig recalled proposing a Claude-assisted interface that could go online that night, with Hui and Lin’s agreement. He emphasized speed: an initial interface built that day could capture interest sooner than one built two weeks later.[8]

Neubig also stressed the research contribution behind the interface: “This would have never succeeded without Xingyao being interested in the project.” Brennan separately credited designer Paul Bloch with an early visual mockup. The April 2025 OpenHands paper credits Wang with project leadership and CodeAct agent implementation, and Brennan with initiating the architecture. Neubig contributed project advice, resources, substantial paper edits, and additional integration and code-review support.[8][10]

In his July 15, 2024 company announcement, Neubig explained why the founders wanted a company: full-time staff and compute could support user-experience work, extensive agent experiments, and deployment in real development processes. His case for openness was also explicit: “everyone should be able to use it and participate in making it better.” In the later founder interview, he argued that shared development could avoid companies duplicating the same work while giving academic researchers and industry users software they could inspect, modify, and deploy locally.[6][8]

In “Don't Sleep on Single-agent Systems”, published September 26, 2024, Neubig organized agent design around three components: the underlying language model, the prompt, and the action space. He argued that separating a software task among specialized agents can introduce rigid workflows, context loss at handoffs, and more code and prompts to maintain. His concrete example was a verifier that needs to locate relevant files but lacks the fault-localization tools assigned to another agent.[9]

Neubig’s proposed alternative combined a capable general-purpose model with tools for writing code, running code, and browsing the web. Those tools let an agent use software already built for human developers. At the time, OpenHands combined instructions by concatenating prompts. He acknowledged the costs: longer prompts consume more money and time, and irrelevant context can distract the model. Retrieving appropriate instructions was a possible alternative, with the best method still an open research question.[9]

His preference was qualified. Neubig endorsed separate agents when they hold privileged information or act for different people. The OpenHands platform also supports delegation. Its April 2025 paper describes agents operating through a chronological event stream of actions, observations, and user feedback, with a Docker runtime providing bash, IPython, and browser interaction. This infrastructure separates agent behavior from action execution and allows users to supply feedback while an agent works.[9][10]

Before OpenHands, Neubig’s academic work included neural-network infrastructure and agent evaluation. These earlier collaborations provide context for his work on tools that researchers can use and environments in which agent capabilities can be tested.[12][11]

Neubig earned his undergraduate degree at the University of Illinois at Urbana-Champaign and his master’s degree and Ph.D. at Kyoto University. He joined Carnegie Mellon’s Language Technologies Institute as an assistant professor in 2016. His research statement identifies a long-term goal of reducing barriers in communication between people and between people and computers, through work including machine translation and natural language understanding.[5][3]

Neubig was the first-listed author of the January 2017 DyNet paper, with Chris Dyer, Yoav Goldberg, and 22 other collaborators. DyNet builds computation graphs as procedural code executes, allowing the network structure to vary with each input. Because a fresh graph is constructed for every training example, the toolkit uses an optimized C++ backend and a lightweight graph representation to keep construction overhead low. Its interfaces support models written idiomatically in C++ or Python.[12]

Neubig later coauthored WebArena with Shuyan Zhou, Frank F. Xu, and other collaborators. First submitted in July 2023, it supplies reproducible, functional websites for shopping, social forums, collaborative software development, and content management. Tasks are evaluated by whether the requested work is completed correctly. The April 2024 revision reported 14.41% end-to-end success for the team’s best tested GPT-4-based baseline, compared with 78.24% for humans—a historical comparison showing the difficulty of that evaluation.[11]

Neubig also makes research and teaching materials publicly available, including papers, advanced NLP course materials, video lectures, and open-source software. His CMU faculty profile lists teaching in large language models and inference algorithms for language modeling, alongside research on model capabilities and applications such as software development.[3]