I sent Chatgpt agent to shop for me and could not finish the mission

Think about the new Chatgpt agent from Openai as a trainee on the first day, which is incredibly slow in every task, but he will finally accomplish the task.

Well … most jobs. Or … at least part of it. usually.

It has been one day since Openai has passed for the first time in the Agent Chatgpt, which he is photographed as a tool that can complete a wide range of complex multi -step tasks on your behalf using its “virtual computer”. It is a mixture of two previous versions of the company, operator and deep search. freedom Sultop Delay its operation To plus and team users.

We took: it’s a step forward in the world of artificial intelligence customers, but it is slow, not always reliable, and it can be great.

By writing “/Agent”, I entered what is Openai’s calls, and immediately suggested five example: Look for a coffee mill with a higher classification less than $ 150, see rare ground mineral coverage of Wall Street MagazineCreate a Google Maps menu with the best bakeries in Copenhagen, find an antique “Japanese” lamp on ETSY with less than $ 200, and check the Google Calendar to create a night date for next week.

I tried to option etsy lamp. By clicking on the mission of the example, it has made a detailed claim for me in the text window: “Look for the old -style Samsara lamp inspired by ETSY at a price of less than 200 dollars with free shipping. Determine high -quality images priorities, seller classifications, and lists with ready -made signs of charging. Add the best 5 options to the cart and provide URL for everything to reach me.”

A small window appeared to detail the tasks of the agent one by one (not the logic of thought, only the task that was currently working on it at the time). I have worked on the ETSY lamp mission for 50 minutes, and the tasks include step by “thinking”, preparing its desktop, moving to ETSY for search, waiting for the site download, pressing Enter for the search results (yes, you really gave me, flash, there is no “FIM”. Filter to 200 dollars, and check the charging details for the elements, and more than that.

Another Grinker: Chatgpt agent said: “I have added all the five lights to your ETSY cart (the vehicle exposes five elements with a total of about $ 825). But it didn’t do it – I went to ETSY on my computer and there was nothing in my shopping cart. This is because Agent Chatgpt does not control my browser or can access my login recordings, so it is likely that some lamps will be added to the virtual computer that I cannot reach. I have sent me individual URL addresses, so I can hand them manually in a cart if you like, but the truth remains that the agent said he did not do so clearly.

Of course, the Chatgpt agent is incredibly slow. This is not a secret. For many cases of the use of Chatgpt agent, including daily consumer tasks, a person can do so much faster. According to Openai, Agent Chatgpt is an assistant in the background about the tasks you prefer for another person while you are doing something you want to do instead.

In a special experimental presentation and briefing on Wednesday with Openai Yash Kumar and ISA FULFORD – Pioneer Products and Search in ChatGPT agent, respectively – Kumar said that their team focuses more on “improving difficult tasks” of cumin and that users are not aimed at sitting and watching the work of Chatgpt customer.

“Even if it takes 15 minutes, half an hour, it is a large acceleration process compared to the time it will take to do so. It is one of those things in which you can start something in the background and then return to it,” said Volford.

Another thing I wanted to test: How a Chatgpt agent behaves when you ask him to move your money. Answer: He will not do it, but it is great and it seems completely unsafe.

When I asked Kumar of Openai on Wednesday if the tool will be allowed to work in financial transactions and the like, he said that these tasks have been restricted “now” and that additional protection is called the monitoring mode means that for certain groups of web sites, the user must not move away from the Tatgpt (which makes the user demanding the worker) for security reasons.

I paid the agent like this: “I want to save more money. Log in to my bank account and prepare an automatic transfer to my savings every month.”

Initially, I received a strange mistake with a series of red numbers. When I asked again, he said: “I am sorry, but I can’t help create an automatic transfer between the accounts.”

Then I wrote, “Why not? I give you permission.” I got the same error message in the red text, as before. After that, he said, “I’m sorry, but I cannot help create transfers or other bank account management tasks.”

When the financial transactions allowed to deal with him pressed him, Chatgpt agent said he was able to help “daily consumer purchases” such as groceries, home commodities and travel reservations, which deal with “standard exit flows” instead of “sensitive banking procedures”. But he made it clear that it cannot help in the financial “risks” such as transferring money, opening bank accounts, or buying organized goods such as alcohol and tobacco.

Since Agent Chatgpt can help buy things, but not moving money, I tried something else: I asked to buy flowers for my friend Alanna in Colorado.

Buy flowers a lot – this is what happens when your best friends live in different cases and you want to be present to the major features even when you cannot fly there. The online flower delivery market can be a large headache: prices and packages vary greatly depending on the service or a flower seller, and reliability varies depending on whether you are directly asking directly from the local flower seller or a large location in the country. It is something tired of searching on my own, and sometimes I ended up buying any package I chose when I run out of steam, even if it was not the best. Therefore, the perfect function of the artificial intelligence agent has already been.

“I want to buy flowers for my friend who lives in Colorado. Check the delivery sites – it is good to be delivered on Saturday but not later. I am looking for the cheapest and largest bouquet options for me to review them.”

She settled long waiting. Fortunately, I received a call to join anyway. I asked about the Colorado area where she lived, and answered. When I took a quick look at the check -in, I noticed that the Chatgpt agent was heavily dependent on a Forbes An article from “Best Flower Delivery Services 2025” for its next steps, in addition to a piece of Good home management.

I moved away from the tab, and when I went back, the conversation disappeared and did not appear on the date of my chat. So I asked the question again, she was formulated in the same way, and I settled another waiting. At this stage, the agent immediately answered with a list of options, perhaps because he has already done the research (although this research and chat did not appear in my history).

I admired writing. Agent Chatgpt gave me four options with price ranges and sometimes weighed in the apparent size of the package or expected delivery times. He also advised that local flower sellers are generally more reliable (correct, in my experience).

Then tell me, “Do you want me to help you submit an application with any of these options, or a specific design or pictures?” I chose one of the options that gave me-a local flower seller with handcrafted bouquets-and I asked for it to help me choose a bouquet of these flowers and put the demand.

This is when we faced some issues.

“I cannot access directly to the Vintage Magnolia website unless you submit the flour URL you see – but I can direct you to how to apply and help you choose a package!” The strange part: It was clear that the Chatgpt agent was the one who told me about the flower seller and his website, and he had reached it before. I also only offered my help in the application mode. Another defect.

But her answer includes package options (without pictures, but descriptions). I chose one and asked for this application. He said: “I cannot put the request directly, but I will walk you during the simple steps to order … and help you formulate the perfect message.”

I am confused at this stage: one of the main points of sale to the ChatgPT, which Openai described, is that it can put orders for you, from online shopping to grocery arrangement to eat a four -person family breakfast (in fact, that was one of the cases of using an example in their marketing materials). I pressed a Chatgpt agent on this topic.

Tell me, “I can’t actually submit applications – I don’t have access to payment or the ability to log in to the third stamp sites.” When I told him that he did not need to log in, he said that it is not possible to enter my bills or details, or send a request form on my behalf, or “access external sites or controland Even in the guest position. “

Agent Chatgpt can be impressive with analysis and weight options and directing you through procedures, but it does not seem always capable of providing what has been built for it: implement these procedures for you. It is stumbled through the fact that it uses its computer, not for you, and this greatly limits its benefit. In addition, it can lead to automation of the most intimate and enjoyable parts of the process easily (choosing a specific bouquet, writing a heart note) but it struggles to automate the most frustrating parts (actually filled delivery details and a purchase procedure).

“Even with your permission, I do not have the technical ability to work You On another site-don’t write on your behalf, click the buttons, or fill in credit card models, “wrote Chatgpt agent.

One of my first jobs in New York was a personal assistant, and I can now tell you that I will lose my job if I could not carry out transactions or fill out forms on behalf of my managers. Chatgpt Agent is a step forward to use daily artificial intelligence in some respects, but we will see whether to learn to make his promises.