How to Use the New ChatGPT Agent If You Trust It
The era of AI agents has arrived: ChatGPT is no longer just a tool for answering your queries, giving confident but often completely wrong answers synthesized based on massive amounts of data obtained from other sources . Now it will connect to your applications and perform real actions: book flights, look up prices, check the calendar, create slideshows and much more.
The new service is called ChatGPT Agent, and it essentially gives an AI bot its own virtual computer inside your conversations. OpenAI promises that the bot will “seamlessly switch between reasoning and acting, executing complex workflows from start to finish, completely following your instructions” — which sounds ambitious, and perhaps a little scary.
If you have any of ChatGPT’s paid plans (starting at $20/month), you can try out the Agent right now. I decided to test it on a few fictitious projects (I’m not ready to trust it with anything real yet). You can launch the ChatGPT Agent in the web app by clicking the “+ ” button to the left of the hint field and selecting the “Agent” mode .
How does agent mode work?
When you switch to agent mode, nothing special happens: you are simply asked to describe the task you want ChatGPT to perform. There are no instructions about your request. Some prompts appear on the screen, from a news summary to ordering groceries.
Once you have decided what you want ChatGPT to do, it can ask you follow-up questions to clarify. The interface is not much different from a normal conversation with an AI bot. The difference is that there is an embedded window that gives you a general idea of what ChatGPT is doing on its virtual computer.
It’s not a direct data feed, but ChatGPT will tell you what it’s doing and display graphs for each action. At any point, you can scroll back or take control of the ChatGPT computer — and you’ll see exactly what ChatGPT is doing, just as if you were connecting to another computer remotely.
You can also switch to what’s called an active mode, where you simply see a ticker with information about ChatGPT’s actions, without any visual effects. You can also stop the agent at any time if you feel that they are going off course or doing something undesirable. It only takes a couple of clicks.
When the ChatGPT Agent has completed everything you asked it to do, you will receive a summary and report. At the end of the final response, as is typical in ChatGPT conversations, you will also see a list of sources. You will be able to ask follow-up questions if necessary.
Overall, the Agent works well, although it does take time: like the Deep Research tool , you’ll likely have to set it up first and then spend some time doing something else. This means you won’t be able to monitor and control ChatGPT’s every move, so you’ll have to decide how much you trust it.
How My Experiment with ChatGPT Agent Went
The first task I asked ChatGPT to do was to plan a birthday party for me: I told it my age, the type of party (quiet, low-key), the location (a small room next to a bar), and the possible dates I was considering. I also asked the AI to come up with some invitations.
Overall, the bot did a pretty good job. It identified local venues that I would have chosen myself, although it did have some issues getting booking information (opening PDFs from the internet didn’t seem to work). The invitation design and copy were nice, if a bit formulaic, and the final report included a handy comparison chart to help me choose a party venue, as well as contact information for booking.
For my next experiment, I tried to get ChatGPT Agent to create a nicely formatted spreadsheet with all iPhone release dates — that would really help me in my work and save time. The real plus was that ChatGPT did a good job of identifying reliable sources: Wikipedia, Apple press releases, and sites like MacRumors.
As far as I could tell, the resulting spreadsheet looked perfectly accurate and was in Excel format. I couldn’t get the nice formatting I wanted, and the source column wasn’t entirely clear, but all the key data was there. It took a fair amount of time to compile, and I think I could have done it myself in the same amount of time (though I was free to do other things while ChatGPT was running).
I am impressed with how user-friendly and functional ChatGPT Agent is. It is not perfect, but in most cases it does the right thing and switches between tasks successfully. The agent is transparent enough that you can always take control if necessary.
Personally, however, I prefer to do these tasks myself. I’m too worried about ChatGPT Agent making a mistake, missing a detail, or not understanding a nuance to rely on it entirely. Your own threshold for such problems may be different, and I suspect many users will overlook minor issues because Agent can save them time.
Disclosure: Lifehacker’s parent company, Ziff Davis, filed a lawsuit against OpenAI in April, alleging that it infringed Ziff Davis’ copyrights in the training and operation of its AI systems.