Homepage | Github | Arxiv | PDF
This project explores a novel approach to web agents by enabling them to use APIs in addition to traditional web-browsing techniques. By leveraging API calls, agents can perform tasks more efficiently and accurately, especially on websites with comprehensive API support.
- API-Based Agent: The API-based agent leverages application programming interfaces (APIs) to interact directly with web services, bypassing traditional web-browsing actions like simulated clicks.
- Hybrid Agent: a agent that combines the power of API-Based Agent and traditional Web-Based Agent, capable of interleaving API calls and Web Browsing.
- Real-World Web Task Evaluation and Analysis: On WebArena, a real-world web task benchmark, our hybrid agent achieve sota performance among task-agnostic models.
We employed OpenHands as our primary evaluation framework for developing and testing our agents, as it is an open-source platform designed for creating AI agents that interact with both software and web environments. Thus, the setting up procedures follow that of OpenHands.
To setup, follow the instructions in SETUP.md.
To run experiments, do:
bash evaluation/webarena/scripts/run_infer.shIn evaluation/webarena/scripts/run_infer.sh, you could change start_task_id and eval-n-limit, where start_task_id is the task_id you would like to start with in WebArena. This also represent the specific website from WebArena you are evaluating. For example, the task with task_id=0 is a shopping admin task, so if you specify start_task_id=0, then it will run shopping admin tasks; the task with task_id=10 is a shopping admin task, so if you specify start_task_id=10, then it will run map tasks.
eval-n-limit is the number of instances to evaluate on starting from start_task_id.
@article{song2024browsingapibasedwebagents,
title={Beyond Browsing: API-Based Web Agents},
author={Yueqi Song and Frank Xu and Shuyan Zhou and Graham Neubig},
journal={arXiv preprint arXiv:2410.16464},
year={2024},
url={https://arxiv.org/abs/2410.16464}
}