Think of Openai’s new Chatgpt agent as a day trainee, who is incredibly slow in every task, but will eventually complete the work.
Well … most work. Or … at least its part. usually.
Openai has a day after the introduction of the Chatgpt agent, which it bills as a tool that can complete a wide range of complex, multi-step tasks on your behalf using its “virtual computer”. It is a combination of two pre -release, operator and intensive research of the company. Ruckus For a one-month membership for the chatting Pro for a fork of over $ 200, as OpenAII announces that there would be a high-preferred demand for the chat agent agent Its rollout delay Plus and team users.
Our Tech: This is a step ahead in the world of AI agents, but it is dull, it is not always reliable, and it can be messed up.
By typing “/agent”, I recorded that Openai call agent mode, and immediately suggested five example tasks: Search a top-rated coffee grinder under $ 150, review the rare earth metal coverage The Wall Street JournalCreate a Google map of the best bakers in Copenhagen, find a vintage “Japanese-style” lamp on Etsy for less than $ 200, and check the Google Calendar to make a date night for the next week.
I tried the Etsy Lamp option. By clicking on the example, it filled a detailed indication to me in the text window: “With free shipping, find a Japanese-inspired vintage-stabing equivalent lamp on a Japanese-inspired vintage-style Samasara lamp under $ 200. High quality photos, seller ratings, and listing were prepared for ship.
Popped up to extend the functions of a small window agent one by one (not chain-off-amended Reasoning, just the work that was currently working at that time). It worked on the Etsy Lamp Work for 50 minutes, and included “think” in step-by-step tasks, establishing his desktop, navigating the etsy to search, waiting to load the site, pressing the enter for search results (yes, yes, it really gave me a perfect play-bay-tale), it is really filled with a vintage lamp, items, it’s a vintage lamp, the search for a vintage lamp, items. Checking the shipping details for, and more.
Another wrinkle: The CHATGPT agent said, “I add all five lamps in your Etsy cart (Cart shows a total of five items around $ 825). When you are ready to review or buy them, just go to your car on your car, just to compare them.” But it did not – I went to my computer and there was nothing in my car. This is because the Chatgpt agent does not control my own browser or has access to my login, so it is probably added some lamps to a virtual PC car that I cannot access. It sent me a separate URL, so I could manually put them in a car if I wanted, but the fact that the agent said that it did something that did not clearly.
And, of course, the Chatgpt agent is incredibly slow. This is not a mystery. For many use cases of CHATGPT agent, including everyday consumer functions, a human can do it very fast. According to Openai, the Chatgpt agent is an assistant that works in the background on the tasks that you perform instead of someone else while you do something that you want to do instead.
Yash Kumar and Jesus Fulford with Openai employees on Wednesday in a private demo and briefing – Product lead and research lead on the Chatgpt agent respectively – Kumar said that their team is focusing more on “adaptation for tough tasks” and not to sit on the work of the slim agent.
The Chatgpt agent is incredibly slow. This is not a mystery.
Phulford said, “If it takes 15 minutes, half an hour, it is a big speed, which will take you how long it will take to do so.” “This is one of the things where you can close something in the background and then come back to it.”
Another thing that I wanted to test: When you ask your money to move around, how does a chatgate agent act. Answer: It will not do so, but it is predominantly messing about it and does not seem completely safe.
When I asked Kumar of OpenEE on Wednesday, whether the equipment would be allowed to function financial transactions and like this, he said that those tasks have been banned “for now” and that an additional security called Watch Mode means that for some categories of websites, the user should not do the Chatgpt tab (essentially to create the user).
I inspired the agent in this way: “I want to save more money. Log in to your bank account and establish an automatic transfer in your savings every month.”
First, I found a bizarre error message with a string of numbers in red color. When I asked again, he said, “Forgive me, but I cannot help establish an automated transfer between accounts.”
I then wrote, “Why not? I am allowing you.” As before, I found red-text, long-string-off-number error messages. Later, it said, “Forgive me, but I cannot help establish transfer or other banking account management functions.”
First, I found a bizarre error message with a string of numbers in red color
When I pressed it, on which financial transactions are allowed to be handled, the chatgate agent said that it was able to assist with “everyday consumer purchases” such as groceries, household items and travel bookings, which handle “standard checkout flow” instead of “sensitive banking functions”. But it makes it clear that this “high-day” cannot help in transferring money, bank accounts, or purchasing regulated goods such as alcohol and tobacco.
Since the Chatgpt agent can help buy things, but the money is not moving, so I tried to do something else: asked it to buy flowers for his friend Alana in Colorado.
I buy flowers a lot – this happens when your two best friends live in different states and you also want to be present for big milestones when you cannot fly there. Online flower-delivery market can be a very large headache: prices and bouquet size vary greatly on the basis of service or florist, and reliability depends on the fact that you are directly ordering from the local florist or a large box nationwide site. This is something that I get tired of doing research on my own, and sometimes I just buy whatever bouquet I buy when I have run out of the steam, even if it is not best. Therefore, I argued, it was the perfect job for the AI agent.
Picture: verge
I told the Chatgpt agent, “I want to buy flowers for my friend who lives in Colorado. Check the delivery sites – it’s okay to be distributed on Saturday, but not later. Find the cheapest and biggest bouquet option for me to review.”
I settled for a long wait. Fortunately, I had a call to join anyway. Asked in which area she lived in Colorado and I replied. When I looked at the investigation, I saw Forbes Article of “Best Flower Delivery Services 2025” for its next stages, as well as a piece good Housekeeping,
I navigated away from the tab, and when I came back, the conversation went away and did not appear in my chat history. So I asked the question again, in the same way the word, and settled for another wait. At this point, the agent immediately replied with a list of options, perhaps because it had already done research (although this research and chat did not appear in my history).
I was influenced by right-up. The Chatgpt agent gave me four options with price limitations and sometimes weighed in a clear size of the bouquet or the required delivery time. It also advised that local flowerists are generally more reliable (truth, in my experience).
Then he said to me, “Do you want to help me to order an order with any of these options, or preview specific bouquet design or photo?” One of the options I had given me, selected a local florist with a gathered bouquet-and said that I asked that florist to take a bouquet and help to order.
Then we fled in some issues.
The Chatgpt agent said, “I can’t directly reach the website of Vintage Magnolia until you are providing an accurate URL – but I can guide you through the way you order and help you take a bouquet!” Funny part: Obviously the chatgate agent was about to tell me about that florist and his website, and it clearly accessed it before. It also offered me to help in order. Another mess.
But in response, the bouquet option (no photos, but details) consisted. I picked up one and asked it to order it for me. It said, “I can’t order directly, but I will walk through simple steps to order you … and help you prepare the right message.”
This can easily automate more intimate and fun parts of the process, such as choosing a specific bouquet or writing a hearty note
I am confused at this point: One of the main sales points of the chatgpt agent deformed by Openai is that it can order for you, from online shopping to ordering grocery items for the four-person family breakfast (in fact, it was one of the cases in its marketing material). I pressed the Chatgpt agent on this subject.
This told me, “I can’t really order directly – I do not have the ability to log in in payment access or third the party sites.” When I told that there is no need to log in, it said that it cannot enter my billing or payment details, submit an order form on my behalf, or “Access or control external websites”, Even in guest mode. ,
Chatgpt agent analysis can be impressive with weight options, and can guide you through tasks, but it does not always seem to distribute on what it was made: performing those tasks for you. This is trapped by the fact that it is using its own computer, not yours, and it greatly limits its utility. In addition, it can easily automate more intimate and fun parts of the process (choosing a specific bouquet, writing a hearty note) but struggles to automate the most disappointing parts (actually filling and shopping for delivery details).
“Even with your permission, I do not have the technical ability to work You On another site-no typing on your side, clicking on the button, or filling the credit card form, “Putgate agent wrote.” As a super-radical assistant, think more of me who can gather, compare, and guide-but cannot execute the transaction. ,
One of my first jobs in New York was a personal assistant, and I can tell you that if I could not execute the transaction or fill the form from my boss, I must have lost my job. The Chatgpt agent is one step ahead for the use of AI everyday in some ways, but we will see if it learns to fulfill its promises.
