Hey there, 

OpenAI just shut down ChatGPT Atlas earlier this month and folded its agentic browsing capabilities directly into ChatGPT instead. This new functionality was added to the ChatGPT Work mode and allows it to navigate websites and complete tasks while signed in as you. ChatGPT has effectively shifted from simply telling you what to do to actually going and doing it. This week we’ll look at what ChatGPT Work is, how signed-in browsing works, and its limitations and risks. 

What is ChatGPT Work, and what can it do?

ChatGPT Work is a new mode that's designed for your longer and more complex multi-step tasks. Now instead of ChatGPT explaining what you need to do, it can go and act for you based on your instructions. This means that it will take in your end goal, plan the steps, run them, and then check in and share progress as it works. ChatGPT Work can create all kinds of different outputs. These can be as simple as documents, spreadsheets, and presentations, or as complex as full websites. The Sites feature that we looked at last week was a feature within Work.

ChatGPT Work can do more than just create though. It can act as a second pair of eyes by watching a webpage for changes. Work is also able to run scheduled tasks which allows you to set up automations with a variety of triggers. You can also connect ChatGPT Work to the apps you use everyday. These range from work related apps like Google Drive and GitHub to your personal apps like Apple Music or the Target app. Once signed in, ChatGPT Work can work within these sites for you, writing and pushing code, creating docs, and buying groceries. 

How the signed-in browsing works 

ChatGPT Work utilizes its own separate browser that doesn’t touch your normal tabs, history, or passwords. It can navigate through and interact with websites and you can watch it happen live. If and when it needs to sign-in to an account, you can manually enter the details on the webpage. Crucially, you should not share your passwords with ChatGPT through the chat. By entering the password manually, you are ensuring that ChatGPT Work can’t see or remember it. To avoid forcing you to constantly input your passwords, Work will hold onto your sign-in sessions for later tasks. Before completing certain actions, such as submitting forms, making a purchase, or changing any settings, ChatGPT Work will stop and ask for approval. 

How to use it

Before we jump into how to use it, we need to cover who can access it. Unfortunately ChatGPT Work is a paid feature which means that only Plus, Pro, Business, Enterprise and Edu users can work with it. Once you’re in, getting started is easy. It starts off the same way as a traditional chatbot where you describe the outcome you want. From there things look a bit different as you’ll need to grant access to websites and sign-in when needed. Once the task is completed, you can see the output and determine if it's acceptable. 

As you use ChatGPT Work, you’ll quickly see the importance of providing relevant context. This means that you describe what you want (for example, finding an appointment at the DMV) and give relevant details (location, purpose of visit). It’s also essential to include limits on what Work should do (do not book any appointments yet). Finally you’ll need to make it clear what the output should look like. In this DMV example, we already did this by saying we want to find an appointment so now ChatGPT Work has an endpoint to work towards. 

Of course this loop works for finding and booking flights, as you can see from the example above, but it doesn't stop there. You could connect to Zillow to help find homes and apartments that meet your needs or find jobs that interest you. ChatGPT Work can also be used to create extensive competitor reports or research new tools for your team. The agentic nature of ChatGPT Work means it can complete a massive range of different tasks and help you get more work done faster. 

Limits, privacy, and staying in control 

It's important to understand the limits and risks of using tools like ChatGPT Work, especially the agentic browsing feature. A common roadblock you’ll hit are CAPTCHAs meant to stop automated browsers. While these AI checks pause the browsing, they don’t actually stop it. While on one website, it paused and asked if it should complete the check and keep going. It did without any issues and moved to complete the task. It’s important to note that I didn’t extensively test this ability and that doing this likely breaks the Terms of Service for most websites as the point of CAPTCHA is to keep automated tools out. 

Like with all AI tools, there is the potential that ChatGPT will misinterpret your instructions so it's important to keep tabs on the progress as it works. This is most likely when you give vague instructions so always try to thoroughly explain what you’re looking for. When it comes to privacy, Work isn’t using your usual browser and will instead use its own browser with unique tabs, and history. As mentioned above, ChatGPT Work can’t see your passwords and instead uses the secure sign-in feature. This helps keep your passwords safe while browsing. 

The most important safety concern to remember is that not all webpages should be trusted. A webpage can have hidden instructions that try to trick your agent into doing harmful tasks or sharing information it shouldn’t. Just because ChatGPT Work can access a site doesn’t mean that the site is actually safe. My advice here is to start with simple, low risk tasks to build up your understanding and comfort with the tool before expanding to tasks that require Work to submit, purchase, or book anything.