[Paper] ♦ [Benchmark Contents] ♦ [Getting Started] ♦ [Live Demo] ♦ [BrowserGym] ♦ [Citing This Work]

WorkArena is a suite of browser-based tasks tailored to gauge web agents' effectiveness in supporting routine tasks for knowledge workers. By harnessing the ubiquitous ServiceNow platform, this benchmark will be instrumental in assessing the widespread state of such automations in modern knowledge work environments.

WorkArena is included in BrowserGym, a conversational gym environment for the evaluation of web agents.

At the moment, WorkArena includes 19,912 unique instances drawn from 33 tasks that cover the main components of the ServiceNow user interface. The following videos show an agent built on GPT-4-vision interacting with every such component. As emphasized by our results, this benchmark is not solved and thus, the performance of the agent is not always on point.

Goal: The agent must search for specific information in the company knowledge base.

The agent interacts with the user via BrowserGym's conversational interface.

Goal: The agent must fill a complex form with specific values for each field.

Goal: The agent must order items with specific configurations from the company's service catalog.

Goal: The agent must filter a list according to some specifications.

In this example, the agent struggles to manipulate the UI and fails to create the filter.

Goal: The agent must navigate to a specific application using the main menu.

Goal: The agent must answer a question that requires reading charts and (optionally) performing simple reasoning over them.

Note: For demonstration purposes, a human is controlling the cursor since this is a pure retrieval task

To setup WorkArena, you will need to get your own ServiceNow instance, install our Python package, and upload some data to your instance. Follow the steps below to achieve this.

1. Go to https://developer.servicenow.com/ and create an account.
2. Click on Request an instance and select the Washington release (initializing the instance will take a few minutes)
3. Once the instance is ready, you should see your instance URL and credentials. If not, click Return to the Developer Portal, then navigate to Manage instance password and click Reset instance password.
4. You should now see your URL and credentials. Based on this information, set the following environment variables:
   • SNOW_INSTANCE_URL: The URL of your ServiceNow developer instance
   • SNOW_INSTANCE_UNAME: The username, should be "admin"
   • SNOW_INSTANCE_PWD: The password, make sure you place the value in quotes "" and be mindful of escaping special shell characters. Running echo $SNOW_INSTANCE_PWD should print the correct password.
5. Log into your instance via a browser using the admin credentials. Close any popup that appears on the main screen (e.g., agreeing to analytics).

Warning: Feel free to look around the platform, but please make sure you revert any changes (e.g., changes to list views, pinning some menus, etc.) as these changes will be persistent and affect the benchmarking process.

Run the following command to install WorkArena in the BrowswerGym environment:

pip install browsergym-workarena

Then, run this command in a terminal to upload the benchmark data to your ServiceNow instance:

workarena-install

Finally, install Playwright:

playwright install

Your installation is now complete! 🎉

Run this code to see WorkArena in action.

Note: the following example executes WorkArena's oracle (cheat) function to solve each task. To evaluate an agent, calls to env.step() must be used instead.

import random

from browsergym.core.env import BrowserEnv
from browsergym.workarena import ALL_WORKARENA_TASKS
from time import sleep


random.shuffle(ALL_WORKARENA_TASKS)
for task in ALL_WORKARENA_TASKS:
    print("Task:", task)

    # Instantiate a new environment
    env = BrowserEnv(task_entrypoint=task,
                    headless=False)
    env.reset()

    # Cheat functions use Playwright to automatically solve the task
    env.chat.add_message(role="assistant", msg="On it. Please wait...")
    cheat_messages = []
    env.task.cheat(env.page, cheat_messages)

    # Send cheat messages to chat
    for cheat_msg in cheat_messages:
        env.chat.add_message(role=cheat_msg["role"], msg=cheat_msg["message"])

    # Post solution to chat
    env.chat.add_message(role="assistant", msg="I'm done!")

    # Validate the solution
    reward, stop, message, info = env.task.validate(env.page, cheat_messages)
    if reward == 1:
        env.chat.add_message(role="user", msg="Yes, that works. Thanks!")
    else:
        env.chat.add_message(role="user", msg=f"No, that doesn't work. {info.get('message', '')}")

    sleep(3)
    env.close()

Please use the following BibTeX to cite our work:

@misc{workarena2024,
      title={WorkArena: How Capable Are Web Agents at Solving Common Knowledge Work Tasks?}, 
      author={Alexandre Drouin and Maxime Gasse and Massimo Caccia and Issam H. Laradji and Manuel Del Verme and Tom Marty and Léo Boisvert and Megh Thakkar and Quentin Cappart and David Vazquez and Nicolas Chapados and Alexandre Lacoste},
      year={2024},
      eprint={2403.07718},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}