Holy Mary mother of everything! Chat GPT AI agent is a game changer!

Hi, Is it free? or not. and can it work with, say, emailing a friend?

i’ve tried creating websites before with minimal success, so I would really love to know how you get ChatGPT to communicate with wix.

I had access to the ChatGPT Agent before my plan expired, and honestly, at first it was difficult to even log into the portal. I remember trying to use it literally the second day it launched, and I needed extra help to get into the interface. I’m hearing a lot of mixed reviews—some users say it’s accessible, others say it isn’t. When I started using it, I went in with the expectation that I’d be able to book flights, pay the bills, schedule meetings, do a bit of shopping, maybe even order food. But in reality, it wasn’t able to handle most of those things. To be fair, I haven’t tested it extensively, especially since my plan has expired now. I keep meaning to renew it and mess around with it more once I have the free time. From what I experienced, it mostly felt like an upgraded version of the memory feature. That seems to be the main use case right now—handling files like documents, scheduling reminders, creating calendar events. But I’m wondering: can it do anything with media, like videos?
One important thing to remember is that when it comes to ChatGPT Agents, a lot depends on which websites are integrated or onboard. For example, I thought I’d be able to order groceries from Amazon through the agent, but that didn’t happen. Amazon, being the giant that it is, doesn’t really want third-party systems getting involved in their operations. So because of that limitation, those kinds of shopping features aren’t available—at least not with the bigger names like Amazon, Walmart, or Target. That said, I think there’s still potential here. If smaller independent retailers and grocery stores are smart, they could partner or integrate with systems like this and carve out a space. That might be the route forward. I’ve seen people doing some impressive tasks with the Agent—writing code, designing flyers, formatting documents, creating spreadsheets, prepping PowerPoint presentations. I have no idea how they’re doing all that, but it looks like there’s a lot of potential for power users. Still, for me personally, I found it a little underwhelming compared to what I was hoping for. It’s also a bit tricky to get the hang of.

I would love to see a full walkthrough or demo of what this thing can actually do. I’m surprised they haven’t done one yet, but maybe it’s because it’s still early—only been out for a couple of weeks, if I remember right. Ultimately, I think it needs more time to mature and really live up to the hype. But if someone wants to try it now just to get a sense of where it’s starting and compare that to where it’s heading later on, $20 isn’t a bad price—if you can afford it.

I've been playing arount with it for a little while, and I've made it add stuff to my amazon shopping cart, book me movie tickets on bookmyshow, so... Haven't tried booking a flight because I haven't had to yet.
and @JC certainly you can make it email your friend so long as you are ready to give it your email cridencials.
Generally speaking, it's truly a game-changer as far as web accessibility is concerned... Should try it with, say, wordpress or youtube...

One of the problems with current large language models is that they get relatively good at tackling the so-called happy path in coding problems, which is when all the interactions are easily predictable, but fail to tackle even trivial edge cases sometimes, potentially resulting in security problems. Furthermore they are also prone to hallucinate, and this problem has actually been getting worse lately, with consequences like misspelling dependencies that don't really exist, opening a window of opportunity for bad actors to register them and perform supply chain attacks similar to typo squatting for humans. Beyond this there are also code quality problems, in which the AI tends to generate extremely verbose solutions to problems that experienced programmers can solve a lot more efficiently, which makes the generated code unnecessarily much harder to reason about.

All the above combined results in a huge pile of bloated code with lots of technical debt, skyrocketing costs from token usage, and since the time and memory complexity of context windows increases quadratically with their size, it's not even that hard for a medium-sized codebase to hit resource limits so the whole thing is extremely unsustainable. While I think it's perfectly possible to build hybrid models that take as much advantage of existing algorithmic solutions as possible to significantly improve their efficiency, and I have my own theories about them that I will start experimenting with soon, I think that doing so will require a huge paradigm shift that may not happen before the current AI hype bubble pops.

One potential time-bomb issue for this technology is a phenomenon in which training new AI models on the output of other AI is known to lead to model collapse due to a yet not understood increased tendency to hallucinate, which is becoming a problem given the proliferation of AI-generated content on the Internet, and might already be adversely affecting the latest frontier models significantly. This content is often called AI slop mostly because it's easy to generate without providing much in terms of actual value.

If using this requires me to share my login credentials with ChatGPT, then that’s a definite no from me. Even if there’s a way to do it manually, I’d want to know if the process is accessible.

Until those concerns are addressed, I think I’ll be steering clear of it.

Also, once you're logged in to a service through ChatGPT, how do you end that session? Is it as simple as deleting the conversation?

Hey guys has anyone used chat gpt agent on the iphone and if so what has been your experiences?
Is it easier to use it on the computer or can you use it on iphone too.

Hi guys.
So with chat GPT agent what else have people tried to do with it.

Hi guys, so I wanted to let you know that I finally activated a ChatGPT plan to try out the Agent. What I attempted to do was to log into this website to post a comment—similar to the example shown above. Unfortunately, when using the iPhone, that doesn’t seem to be fully possible. I turned on Screen Recognition and was able to confirm that the username and password fields were on the screen, but they weren’t accessible with VoiceOver. If this is a bug or an accessibility oversight, it needs to be reported to OpenAI so it can be addressed as soon as possible.

On the bright side, ChatGPT did successfully manage to navigate to the website and locate the login page, which worked well. I was also able to type my username and password directly into the chat, and the agent was able to enter those details and log me in. That said, from what I’m seeing so far, you have to be extremely specific with your instructions. In some cases, you need to know exactly what you’re looking for in order to get the results you want. For example, I wanted to post a comment on this specific post, but I couldn’t remember the exact name or title. That ended up confusing the model a bit, so maybe websites with a clearer structure or layout might work better. I realized that ChatGPT doesn’t automatically recognize that a post is about itself, which makes sense, but it means you’ll need to be extra clear when giving instructions on sites with dynamic content.
Let me see if I can explain this a little clearer. So imagine you’re on a virtual supermarket website. You decide that for breakfast today, you want a box of Cocoa Puffs, a bottle of Pepsi, and a loaf of bread. Now, on these virtual supermarket shelves, ChatGPT is scanning through categories like “Cereals,” “Beverages,” and “Bakery.” If the Pepsi is sitting in the “Refrigerated Drinks” section or the bread is in “Bakery,” then ChatGPT will likely find those items pretty quickly because it knows where to look and what those categories typically mean. But let’s say there’s another person who owns a completely different website—like Mary, who runs a baking site. She sells chocolate chip cookies. Now you say, “ChatGPT, order me a box of chocolate chip cookies and a sugar-free glazed blackberry doughnut.” If the doughnut section is clearly labeled or easy to access, the model might find it right away. But if Mary filed her cookies under something more abstract like “Mary's Confectionaries” or “Sweet Bites,” ChatGPT might still be able to get there—it’ll just take a bit more time and work. That’s the part I’m trying to highlight. For the model to be most effective, you need to be specific. The reason I couldn’t post my comment on the site was literally because I didn’t remember the title of the post, and I couldn’t recall which section it was under. If you don’t have a good mental layout of the website, it can be much harder for the model to perform the task, even if it gets you in the right general area.

It was able to locate the username and password fields easily because those are common across websites and clearly labeled. ChatGPT understands those elements well—it knows, “This is the login box, and this is where I need to input credentials.” But if something is tucked away under an unusual label or section that isn’t visible on the screen directly, I don’t know how many places the model actually searches before it gives up or times out. Unfortunately, I didn’t get to explore that part much because, like a lot of people are discovering, there’s a time limit. Once you hit it, you’re no longer able to interact with the agent for the rest of the day, and I had already used up my window.
Right now, many of the more advanced features are limited. It looks like you only get 15 minutes per day—or maybe per session—with the browser, though I’m not entirely sure yet. I assumed I’d be able to talk to the agent hands-free in voice mode and have it carry out the tasks for me, but that doesn’t seem to be possible. I noticed that when the task is completed, my phone vibrates and I get a notification—which is a nice touch. It’s definitely a bit slow, but that’s expected given that we’re still in the early stages. If someone were going to do a full review of the product, I imagine they’d need to edit the pause time or task to fit while the model processes everything in the background. Anyway, I couldn’t get it to post the comment, but this is only my first time using it. I’m assuming things will improve in the future as they continue building it out.

Hi,
So when using this on Windows, both through my browser and the desktop app,the virtual browser, the browser you can use to enter your username and password, plus takeover from the agent in general if you need to click something the agent won't do like a Captcha is totally inaccessible. I've tried with JAWS and NVDA, NVDA object navigation OCR, and the JAWS cursors and OCR but nothing works. And it won't go to amazon.co.uk or amazon.com at all, even if I tell it to go to this page without completing any task. There's a checkbox on audiogames.net that it won't click because it's a Captcha, and if I take over from the agent, I can't access the checkbox no matter what JAWS or NVDA commands I try. I mean I could give it my username and password for something, but I'd have to keep changing the password just in case it stores it and my security is compromised. I'd only give it my credentials to log into something if something was really inaccessible, but I'd be changing my password after logging out that's for sure.

I was wondering too about how accessible interactions are, as it is using a VM. SO it seems Stephen is doing tasks that do not require him to interact with the virtual browser?...
As for usage, according to a chatGPT.com page, the Pro plan allows you 400 messages a month. So I guess try to pack those messages?
It is an interesting project for sure and I will keep monitoring it but need some more advances before it can help me with my job.
By the way, Claude has a similar agent but its not been in the news lately.

When I was using the ChatGPT agent this morning, it disconnected, and I couldn’t get it to reconnect again. I’m pretty sure I saw a time and date saying when it would be working again—though I could be totally wrong about that. But the second I saw the message, I instantly assumed the product was limited in some way, kind of like how the advanced voice feature is restricted. A lot of the more advanced features with ChatGPT seem to come with limitations, which makes sense. I mean, with the agent especially, it’s pretty obvious why—many members are trying to use it, and the system needs to keep up and handle all those tasks efficiently. I don’t even think anyone using the free version is going to get access to the agent. If they do , it’s gonna be extremely limited. So if I want to explore more of what it can do, I’m gonna need to play around with it some more when I have the time.

Now, regarding Amazon and shopping—based on what I’ve been reading online, Amazon is not one of the supported shopping websites you can use through the ChatGPT agent. And again, this isn’t that surprising. Amazon has worked hard to become one of the biggest names in online shopping, and the last thing they want is some third-party AI stepping in and acting as a middleman. They’re not going to give that kind of access freely. My thinking is this: smaller businesses, if they’re smart, will absolutely jump on this opportunity. If they can integrate with the agent, lower their prices, maybe offer free delivery or other perks to shoppers—then I could see customers choosing to shop with them instead. This could be a major advantage for smaller vendors looking to grow. As for whether there’s an official list of supported shopping partners, I’m not sure we have this feature as yet, but it certainly seems like the next logical step in the chain of evolution based on current trends.
I haven’t played around with the agent enough to speak definitively on everything. But I do think it depends on what you already know. ChatGPT can browse the internet and get relevant info, sure—but the more you understand about the site you’re trying to use, how it works, and what to look for, the more effective it seems to be. Some tasks are always gonna be easier because they’re direct and straightforward. Others, though, are going to be more obscure or ambiguous—and that’s probably where a lot of the confusion and inconsistency comes in.
I didn’t know that Claude had an agent-style product of its own. I might have to subscribe and check it out. I’ve never subscribed to any of Claude’s plans, and that’s mostly because I’ know the context window—like how many messages you can send in a chat—is limited. Even on the paid plan, I’ve heard it fills up quickly. And instead of starting a new thread when you hit the limit, you only find out when your message doesn’t go through. Another thing I don’t like about Claude is that if I’m typing a message and I accidentally close the app or something interrupts me, the entire message disappears. It’s not like ChatGPT, which keeps the text in the box, so when you reopen it, your content is still there. That’s one of those little actions that makes a big difference.

Don’t get me wrong—Claude gives grounded, logical responses. It's more human than ChatGPT in certain ways. But because of those limitations, I’ve been hesitant to give it a serious try. I’m going to take a closer look and do some research myself. My biggest issue with Claude has always been the censorship and restrictions—it’s more limited than ChatGPT in that sense. They're trying to be that “ethical, moral” AI, but in doing that, they might be missing the mark a bit. Not trying to knock them too hard—they do have a solid product. It just needs a bit of refinement… or loosening up.

First off, I haven't noticed any time limits per session as such. The limitation however is that for plus users, there are 40 chats using agent per month. that's like 40 tasks. Also, the virtual browser, as some of you mentioned, is inaccessible. I guess for it to work, the screenreader providers will have to work with open ai to implement a sollution. That's why as of now, we will have to provide the login cridencials to the agent. What Claud has is Claud compute, which is arkitecturally different from gpt agent. agents generally creates a VM in the cloud, whereas what claud compute does is take over your computer which means it can also access your files etc in the computer.

Hi,
If people want to try Claude Compute, the Guide AI Assistant for Windows uses it as their model.
https://www.guideinteraction.com/
It's about $8 a month at the moment, which I imagine is cheaper than Claude.

Honestly I am wondering if the output, in this case the website was checked by sighted people as well. Sure, AI will tell you that it did what you asked it to do. But even Apple admits that there is at best a 72 % chance that the info AI gives is correct. Or in other words: Do not trust AI to do stuff for you which you can not a) do yourself and b) you are able to verify that the expected outcome has been reached.

I get why the hype regarding AI seems so amazing. But honestly: In most cases it is bloated machine learning which wastes so much water and energy. Creating a Website should not waste galones of water. Thanks to the fairly easy HTML you could write one yourself which would load lightning fast. And as a bonus: You learn how stuff works. If AI put in something that you did not want but everything else was perfect, you would have a hard time just getting this part out. Instead AI will attempt to rewrite the entire thing thanks to your new promt. That might change the complete page.

I am not against AI. But I am against the hype with more and more promises even though the last couple of releases did not work as advertised. Maybe one should take a step back and evaluate how much AI really does do correctly all of the time. Cause computers generaly are pretty good at doing the exact task over and over again without suddonly injecting other stuff no one was asking about. The only benefit I can see is that we have a better way to get pictures described to have an idea what they show. And even that will fail when you ask for detials. When you take a step back and reflect on this you might encounter how the output is currated under miserable working conditions for not a lot of money. In the end human labour is trying to correct for the flaws that are inherent in AI. That should in my opinion not be acceptable.

How can I Even get into the Interface? I asked it to log in to a website and it told me to Type in the credentials Modell in a browser Window. Then there is a windows called Virtual browser. But how exactly can I See the website then? In the Virtual browser window it just Telfs me that I have Control over the Virtual browser and a button called Stop.

The virtual browser on the website isn’t accessible. Yes, ChatGPT can manually type in my username and password for the site, but it can’t complete the CAPTCHA—which is understandable. That part is fine. The real issue is this: if I navigate to a website like American Eagle, Netflix, eBay, or Instacart, and I need to take over manually—whether it’s to read my messages, browse the site, or move around until I’m ready for ChatGPT to assist—I’m stuck. Once I take over on the virtual browser, the screen is taken over by that interface. Using the screen recognition feature, I can tell I’m on the correct website, and I can even tell that the information is somewhere on the screen—but I can’t actually interact with it. That’s a huge, huge accessibility issue. Until this is fixed, the ChatGPT agent is not as useful as it could be to blind users. Yes, I’ve already sent an email to ChatGPT, but I don’t know if or when this will be resolved. Hopefully they fix it, but you never know. ChatGPT can log into Instacart for me, sure—but it can’t let me take over and browse. I can’t explore the interface, go to the produce section, pick out the milk I want, then ask ChatGPT to help finish the checkout. That only works if I can interact with the site—and right now, I can’t, unless I spell out every step, which completely defeats the purpose if I don't know the exact elements that are on the website.

Well then, with all due respect, I guess I must be sophisticatedly stupid—because no matter how hard I try, it never seems to work in my favor. I’m getting 40 messages with the agent. And if I have to waste those messages trying to prompt the agent to do a task that I could easily do on my own through the browser, then that’s a complete waste in my book. I went on Project Gutenberg and had it read the first two chapters of Alice in Wonderland for me, and that worked pretty well. I wish the ChatGPT agent supported voice commands properly. If it worked with the voice mode, I could use the microphone to give instructions directly—but I’m not sure if that’s a feature yet. From what I understand, the advanced voice feature is limited too, so I’m not sure how far I’d even get with that.

Yes, if I get stuck at a CAPTCHA, I’m done. If I need to manually interact with a website for any reason, I have to go back and forth with it—and that takes a lot of time. Maybe it’s different on a computer, I don’t know. But on a phone, it’s pretty slow, and tasks that should take seconds end up dragging for minutes. Sometimes it takes 2–3 minutes to complete a task that shouldn’t take more than 15 seconds. Of course, I know that the product is new, so this is to be expected. For now, at least. I'm not holding this against the developers. This is one aspect I can confidently say will most likely be improved in the future.

Now I’m not saying the agent is completely useless, but the fact remains—it’s not as accessible as it could be. You said it’s all about sophisticated prompting, but I’ve been using ChatGPT since it launched. I’ve seen most versions. I’ve paid for different features when I needed them. I know what I’m doing. So I’m not new to this space—not to ChatGPT, not to Gemini, not to any of these tools. I’ve been as specific as I possibly can. So no, I don’t believe the issue is on my end. The virtual browser isn’t accessible for certain tasks. If I try to take over the browser and interact with the elements myself, nothing happens. That’s a serious accessibility issue. But hey—different strokes for different folks. Plus, I’m not sure if this is only happening to me, but whenever the agent disconnects for whatever reason, I can’t seem to find the option to reconnect within the app itself—I always have to go to the website to reconnect, and it tends to disconnect quite a lot during sessions, so I’m not sure if it’s different on the pro plan or what, but there are a few issues that need to be ironed out.

According to my spouse, it did the website perfectly. Not only that, but it put the comment on Apple this perfectly too. When it comes to things like website design you really need to be specific in your prompting. While you could learn HTML, that’s not gonna help you when it comes to format and look of the website. While it may be functional to us blind users, it may not look visually appealing to the sighted population. If you’re running a business, you do kind of want to appeal to the majority.

You're kinda trying to sell the idea that a non-deterministic service can do a better job than a deterministic language that gives you full control over all visual aspects and is also produced by the aforementioned non-deterministic service, which makes absolutely zero sense. Just because the sight might look good doesn't mean it's not possible to accomplish the same or even much better writing the code yourself, and the fact that the AI itself has to express your intent in that code is irrefutable proof of that.

I'm totally blind and do both user interface design and computer graphics, taking advantage of having lived most of my life with sight as well as the fact that in the end it all boils down to math. You can definitely do at least just as well as the AI, it requires getting creative with your solutions, like investing on a graphics embosser like I'm on the verge of doing, but it's all within the reasonable realm of possibility. Just a couple of weeks ago I designed the logo for the international brand that I am in the process of registering in vector graphics, because I knew exactly what I wanted visually and I knew how to express that mathematically. However since I don't fully trust other people's opinions as they have failed to tell me about important visual details in the past, I really need to feel the visual stuff that I make, hence my plan to invest on a graphics embosser in the near future.

I do think I understand what you're saying, João Santos. The AI is helping me on the other side of all this. I am taking pictures with my phone. I can't exactly tell if an unstaged nature shot, for example, has anything interesting until it is described, and then it is very rare that the picture is well composed, as is.
Instead of having the ability to line up a well composed shot through the viewfinder, which I had experience with when I was sighted, I am having to take on the iOS cropping tool. The AI models I've been using seem to be fairly good at geometry and compositional analysis. So I can ask the AI to describe the contents of the bottom, horizontal eighth of an image, or the contents of any number of grids, building a model in my imagination. I can also ask the AI to describe the image in relation to the intersections of a grid of thirds. It can tell me about leading lines and negative space in an image.
It is horrendously tedious, and can take many hours to achieve what a sighted person could do in a few seconds with a flick of a finger. on the screen because every change has to be loaded into and described by an AI app of some sort or another, after saving the changes, and many times restarting all over from the original image etc.
I think the AI could help analyze the arrangement of a web site being designed by a blind person for visual people after being coded by a person in similar ways, by giving feedback. I want to have some hands on control over what is taking place, rather than the AI generating something that may or may not have weird artifacts or code... An image embosser would be great, but also might be as tedious and much more costly than asking AI. We'd still probably crawl over broken glass to get it done. The audience doesn't suffer for the artist...