Modeling Distinct Human Interaction in Web Agents

Faria Huqc,†, Zora Zhiruo Wangc,†, Zhanqiu Guoc,‡, Venu Arvind Arangarajanc,‡, Tianyue Ouc
Frank Xuc, Shuyan Zhoud, Graham Neubigc Jeffrey P. Bighamc
c Carnegie Mellon University, d Duke University
† Indicates Equal Contribution, ‡ Indicates Equal Contribution
Paper Link Code Huggingface

Overview

🧭

Intervention-Aware Web Agents

Builds collaborative agents that anticipate when people will step in—reducing unnecessary interruptions while preserving user control.

🧪

Real-User Behavioral Data

Grounded in COWCORPUS: hundreds of real collaborative web sessions with interleaved human/agent actions, enabling principled modeling of intervention behavior.

🧠

Style-Conditioned Collaboration

Adapts to different interaction styles (hands-off, hands-on, collaborative, takeover) so agents align with how individual users prefer to supervise and share control.

⚡

Better Timing, Better UX

Predicts high-value decision points and asks at the right moments—improving usability and keeping execution moving when human input isn’t needed.

Data & Performance

Video Demo

Latest Updates

Resources

📄

Research Paper

Read our detailed research paper on CowCorpus and modeling user intervention pattern.

View Paper →
🤗

HuggingFace Models

Explore and use our pre-trained models available on HuggingFace for browser automation tasks.

View on HuggingFace →

Installation & Getting Started

1

Prerequisites

Ensure you have Node.js >= 16 installed on your system.

2

Clone Repository

git clone https://github.com/oaishi/CowPilotv2.git
cd CowPilotv2
3

Install Dependencies

yarn install
yarn add -D typescript @babel/preset-typescript @babel/preset-react
4

Build Extension

yarn start
5

Load in Chrome

  1. Navigate to chrome://extensions/
  2. Toggle Developer mode
  3. Click Load unpacked extension
  4. Select the build folder generated by yarn start

BibTeX

@misc{huq2026modelingdistincthumaninteraction,
      title={Modeling Distinct Human Interaction in Web Agents}, 
      author={Faria Huq and Zora Zhiruo Wang and Zhanqiu Guo and Venu Arvind Arangarajan and Tianyue Ou and Frank Xu and Shuyan Zhou and Graham Neubig and Jeffrey P. Bigham},
      year={2026},
      eprint={2602.17588},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2602.17588}}