is The Verge’s senior AI reporter. An AI beat reporter for more than five years, her work has also appeared in CNBC, MIT Technology Review, Wired UK, and other outlets.
Think of OpenAI’s new ChatGPT Agent as a day-one intern who’s incredibly slow at every task but will eventually get the job done.
Well… most of the job. Or… at least part of it. Usually.
It’s been one day since OpenAI debuted ChatGPT Agent, which it bills as a tool that can complete a wide range of complex, multi-step tasks on your behalf using its own “virtual computer.” It’s a combination of two of the company’s prior releases, Operator and Deep Research. The Verge forked over the $200 for a one-month subscription ChatGPT Pro, since OpenAI announced that higher-than-expected demand for ChatGPT Agent will delay its rollout to Plus and Team users.
Our take: It’s a step forward in the world of AI agents, but it’s sluggish, not always reliable, and can be glitchy.
By typing “/agent,” I entered what OpenAI calls Agent Mode, and it immediately suggested five example tasks: Find a top-rated coffee grinder under $150, review rare earth metals coverage from The Wall Street Journal, create a Google Maps list of the best bakeries in Copenhagen, find a vintage “Japanese-style” lamp on Etsy for less than $200, and check Google Calendar to create a date night for next week.
I tried the Etsy lamp option. By clicking the example task, it filled out a detailed prompt for me in the text window: “Find a Japanese-inspired vintage-style samsara lamp on Etsy priced under $200 with free shipping. Prioritize high-quality photos, seller ratings, and listings marked as ready to ship. Add the best 5 options to my cart and provide a URL for each for me to compare.”
Not quite there. Image: The Verge
A small window popped up to detail the agent’s tasks one-by-one (not the chain-of-thought reasoning, just the task it was currently working on at the time). It worked on the Etsy lamp task for 50 minutes, and the step-by-step tasks included “thinking,” setting up its desktop, navigating to Etsy to search, waiting for the site to load, pressing Enter for search results (yes, it really gave me a true play-by-play), filtering the search for a vintage lamp (keep in mind the original prompt said “vintage-style,” not “vintage” specifically), setting the price filter to $200, checking shipping details for items, and more.
Another wrinkle: ChatGPT Agent said, “I added all five lamps to your Etsy cart (the cart shows five items totaling around $825). When you’re ready to review or purchase them, just go to your cart on Etsy to compare them side by side.” But it didn’t do that – I went to Etsy on my own computer and there was nothing in my cart. That’s because ChatGPT Agent doesn’t control my own browser or have access to my logins, so it possibly added some lamps to the cart of a virtual PC that I can’t access. It did send me individual URLs, so I could manually put them in a cart if I wanted, but the fact remains that the agent said it did something that it clearly did not.
... continue reading