Limitless: An AI Podcast - OpenAI Announces the ChatGPT Agent: Everything You Need to Know

Episode Date: July 23, 2025

In this episode, we explore OpenAI's ChatGPT Agent, an AI assistant designed to enhance productivity by automating daily tasks. We react to its features and discuss its potential impact on finance and e-commerce, as well as the workforce. Tune in for insights into AI's capabilities and ethical implications, and join us next week as we examine the future of AI browsers!------💫 LIMITLESS | SUBSCRIBE & FOLLOWhttps://limitless.bankless.com/https://x.com/LimitlessFT------TIMESTAMPS0:00 Intro1:25 Initial Reactions to the Agent4:04 Testing the Agent's Capabilities7:09 User Experience with the Agent9:16 Enhancements and Limitations11:50 Future of AI and Agents14:26 Financial Applications of the Agent15:52 Everyday Use Cases for the Agent19:53 Impact on Retail and Ecommerce22:47 Security Concerns with AI Agents25:23 The Workforce Transformation28:03 Conclusion and Future Outlook------RESOURCESJosh: https://x.com/Josh_KaleEjaaz:https://x.com/cryptopunk7213------Not financial or tax advice. See our investment disclosures here:https://www.bankless.com/disclosures⁠

So, EJA, you're wealthy. You've just won the lotto. You have all this money. What's one of the first things you do? And I know you're probably going to say a lot of answers, but there's one particular answer that I want you to say, which is you're going to get yourself an assistant, right? To kind of help you do all the things that you don't like to do. They're going to manage your calendar. They're going to order your groceries. They're going to book your flights. They're going to take care of all of the essential annoyances that you just don't want to deal with. Now, what's cool about the news from this week, week is that Open A has released something that essentially commoditizes that part of wealth. It gives you an assistant that allows you to do all of these tedious things that you normally don't like to do, but only for the cost of $20 a month. And it's a product called Agent. And here we'll show up the announcement post that they shared. And what's really interesting about Agent is it kind of builds upon a lot of other things that OpenA has been building, right? They have a browser feature. They have a deep research feature, but they haven't created a fully comprehensive assistant. And that is what I'm assuming Open AI is trying to do with their agent feature. Now, the news dropped last week,

I was watching the live stream. Ijaz, we were kind of chatting throughout the whole thing. You were slightly underwhelmed because it didn't seem as cool as you wanted. I was more like in awe at the user interface. I thought it just looked very pretty more than anything else. And we're kind of seeing on screen the announcement video. What were your first takes when you saw this, the agent from opening AI. Okay, so we were both watching that live stream at the same time, and we were sharing our thoughts. Do you remember that one sentence that I had for you, Josh, which was, if this is another Google Drive connector, another agent that claims to do a ton of things and really it just kind of connects to your Google Drive, I'm going to be really mad. And what are we looking at on our screens

right now, Josh? Tell me what we're listening. That's so funny, because I was reading your message. You just sent this about the first minute of the presentation and maybe 10. 10 minutes later, they show the exact connection that he was describing, which is a Google Drive connector. And what does that mean? For the people who aren't sure why that's not super exciting to you, why is that annoying? So everyone uses email, right? And the most popular email is probably Gmail. And it comes with this suite of different products, right? You've got Google docs where people write up documents at different company jobs. You've got Google Me, people do video conferencing? It's a whole thing, right? And when agents started becoming popular,

Josh, do you remember one of the main claims to fames was that you could now use AI in your Google account, which means that it kind of becomes chat GPT, but it can reference all your documents. It understands all your Excel sheets, your emails, it writes emails for you, whatever, right? But it ended up being more for PR purposes than what the agent actually did. When you actually tried some of these agents, all it kind of did was it felt like a very friction-filled process, very fragmented still, Josh. I still had to tell the AI, no, I meant I want you to reference this document or I want you to talk about this particular thing and I just ended up doing it all myself. So I kind of have a shaded experience with using agents. So when I saw this on the live

stream, I was like, oh God, is this the same thing? And is it the same thing? Do you feel it's the same thing? Or is this a slight enhancement? Because it appears to me like it's different. This time it is a little bit different. Okay. I will humor you. It actually, this time actually is a different. So you mentioned earlier, Josh, that this is a combination of their operator agent, which is, if I'm not mistaken, a computer use agent, right? So if you imagine it is an AI that can look at your desktop or your laptop screen the same way that I'm looking at it right now. And it can scroll with its own cursor. It can open up new tabs. It can open up documents. It can type your keyboard virtually and write a message, right? And it combines this feature with

something called deep research, which you and I have spoken about a lot on this show, Josh, which is kind of like a PhD level researcher for whatever request that you might have. So in theory, that sounded really good. Well, we just got access to this. I'm just going to pull up this tweet so that people have more context here, because I'm sure a lot of our listeners are kind of like asking, you know, can I use this right now? The feature just rolled out to all pro users. So anyone that's on the GPT Pro Plan will get access to this.

And so I was playing around with this for the last couple of days, Josh. And to answer your question, yes, it is actually quite useful. I kind of use some of the presented prompts. So you know kind of like when you haven't typed anything yet, they kind of suggest things that you have to type, Josh. And I got it to organize bouquet of flowers to be delivered to my apartment for my girlfriend. So sweet. That's all I asked it to do.

I was like, you know, off you go. This is not something I want to spend my time doing. So you go and do it for me. Right. Exactly, right? And so I said it, and then I'll tell you my first-hand experience, Josh, which is kind of funny. I just stared at the screen and I watched it create this virtual browser. So for those of you are listening to the audio here, kind of imagine a pop-up appearing and it's like a mini desktop. But all the actions, all the movements, all the screen grabbing and all that kind of stuff is completely autonomous. You're just watching the AI do its thing. So it opened up a browser, Josh, and it started scrolling. through different links from Google, from GPT, of flower websites in my local area. It started using my location tracker to figure out where I was,

so it could figure out kind of like my vague general address, and then figure out what flower shops were near me. It then kind of like started passing through a bunch of different bouquet flowers, all that kind of stuff, kind of figuring out what's seasonal and what's not, what's within kind of a reasonable price point. And then it gave me a bunch of suggestions. And it said, hey, okay, these are,

some pretty good suggestions that I've got for you, but I kind of want to see whether I'm on track, whether I'm kind of hitting your vibe. Can you give me some more information to let me know? And so what I basically did, Josh, was I reviewed what it had for me. I scroll through some of the examples that it had, and I said, yeah, I actually kind of like this.

And options one and four are kind of like really my vibe. And it said, okay, say no more. And it went away and it started like loading up the checkout sides and I could see the way it thought, Josh, which was like one of my biggest takeaways was, it kind of gave me comfort being able to watch it do its own thing. So it took about 10 minutes in this entire process, and by the end of it,

I had a pretty good idea of which bouquet I wanted, and it just needed me to fill out some wallet details, and off I went. So my initial take was, you've just saved me a lot of time. You did that in 10 minutes, where I could have just technically looked away from the screen and got on with my own thing, right, and then come back to it.

The other takeaway that I have, And I'm going to take the other side of this, right? Is it still required me to do things, Josh. It didn't know me entirely. And maybe I'm being too hard on the AI. But ideally I want an AI agent that just kind of like knows what I want and gets it done for me. Right. And like has access to my wallet and pays for me.

We're not quite there yet. But this is a noticeable step change from what we had before. Okay. So good, not great. It'll get you most of the way there, but it won't quite finish the entire thing. And it was interesting hearing. the process because during the stream I was listening to how it works. And you described the combination between operator and deep research, how they're complementary. I kind of want to unpack that a little bit

because operator was the browser-based feature. It could control a browser. So now you have this tool that control a browser, but it can't really read long articles. It's not good at reading long articles. It doesn't have a lot of context. It's not that good at things that deep research was good at. Deep research is good at reading huge amounts of information, compressing it down, coming up with links, and feeding that information back to the browser. So we have this very complementary thing where deep research can't interact with visual web pages, but it can understand a lot of context.

An operator can interact with visuals, but it doesn't have the ability to read a lot. And that creates this complementary tool set that we're seeing on screen right now, which is basically what OpenAic calls the agent. It's given a tool set, so the agent spins up a virtual machine, so you can imagine when you prompt into query,

it created this virtual computer that started talking through how to get the bouquet of flowers. It has a text-based browser tool, it has an image-based browser tool, and it has access to its own terminal to create complementary things. So let's say you were, for example, doing a spreadsheet and you wanted to see what the costs of all the bouquet of flowers are throughout the neighborhood. It can actually do that because it has all these complementary tool sets. So while I can do all of this, well, it's not completely finished, like you said. You still have to kind of nudge it through. So an example that they used in the live stream was trying to get a wedding out.

for a wedding. And it'll get you kind of close, but then it doesn't quite know the sizing. It doesn't quite know your like style preferences. And there's there's more context that it needs to complete this. Yeah, exactly. Yes, it's lacking the context and it's lacking the ability to click pay. So we're seeing here in the example, it's stuck on the pay screen because it doesn't have your credit card information. And they just don't have that capability stored yet. So I would imagine this is probably just, I mean, this is version one. It will get better. It will have. have the context, it will get your credit card info. One of the things I loved, and you described this briefly, is when it loads up this browser window, you can actually replay it. So, Ejaz, you were at

your browser for 10 minutes, kind of watching it do its thing. But in the case that you stepped away and you left, you could come back and you could actually rewatch all the steps that it took as if you were watching a YouTube video, which is something that I really liked a lot about this. I think of all the things that I liked the most, it was the interface. It was very pretty. It was very easy to understand. And it's really funny, the example, they were able to just kind of pan through like it was a YouTube video, see where it made the searches, see what it was looking for, walk through that step process to see how exactly it arrived at the answer. And that was pretty neat. So, I mean, I haven't gotten a chance to try yet. My account still does not have enabled. Please give it to me. But based on your experience, it seems like it's good, not great. And I think that's a step in the right direction because we went from pretty mediocre to good. And that's a big jump in the right direction. Yeah. So let me unpack my thoughts a little more for you, Josh. Basically, I think it's good for people that have high agency. And what I mean by that is people who know what exactly they need to get done. And maybe what they need to get done requires general access to a browser and a few different tools will get a ton of value out of this agent product.

I'm not saying it's useless. It's definitely way more useful than any other agent that's been released before. but for the general public who is so used to I feel like at this point Josh magic when it comes to the internet it's like you know when you open up an app and you're like hua like snaps of photos that disappear

after six seconds that's so novel right back in the day when Snapchat was created you know they want like that kind of like easy to understand kind of like basic level thinking I don't need to think too much it just wows me we're not there yet with agents but I have a feeling that when you combine

agent with deep research as they did today, but with the context of memory, which as you and I have spoken about many times, that open air has the kind of like forefront leadership on, when you combine all those three things together, then you're going to see that magical experience where all you have to do, well, maybe you don't even have to prompt it, Josh. Maybe you just kind of open your screen and the agent is like, hey, I kind of ordered your coffee for you already. It's actually outside the career. I've been tracking him is arriving. And I've like ordered your Tesla Robotaxi to arrive. So, you know, you'll be ready to go and leave at 7.30m. or whatever it is to go to the office. That's when people are going to start being like,

I will pay hundreds of dollars for this thing. And I won't even think twice. Yes. So we have great examples of people using this so far. I see you have one pulled up right here. I want you to walk me through how you can actually use this thing. Because a lot of these things are kind of constrained by your own creativity. I do a lot of stuff, but I'm not quite sure what I'm able to optimize for. So what are some people using this staff for currently? So in this example, this person is using it to shop, which isn't exactly something that I can relate to. I don't do a lot of online shopping, but I have a girlfriend who does. And I know that there are many different occasions that she needs to kind of cater and figure out what to buy for, and there's so many sites that she needs to browse.

So in this example over here, you can see like, you know, she puts in this problem and we pause it. She goes, find a beige trench coat for under $500. Any website is fine, must be under $500. including shipping, must have a belt and double-breasted buttons, optional to have a hood but must be detachable. Basically, all these different specificities that she needs for this comic, right? And kind of reading this kind of gives me a little anxiety because I'm like, oh my God, like I do not have the energy to search for something like this myself. And you see the agent responding to me like, understood,

I'll start searching for the beige coat under 500 bucks and off we go, right? So that's happening. It's doing its own desktop thing and we see it kind of unfolding. it's going through Amazon. It's kind of like going through a bunch of different options, looking at different colors, making sure it fits all the different criteria. And in the meantime,

she's timing up another prompt. And she's running up another prompt. So she's basically running like a bunch of different prompts for like her own chat GPT discussions. And what she's demonstrating here is she can go back and forth between Windows, Josh. And she can like basically multiply her time and effort over a ton of different tasks

while she has this agent doing it for her in the book. background. And if we get to the end of it, as we see over here, we see that it's pulled up a bunch of options. I think it's primarily used Amazon as its main kind of retailer. And then she ends up kind of like making a decision around what kind of coat that she wants to buy. So I thought that was like a really generally accessible or publicly accessible example that anyone can get into. I thought this was pretty fun. So this is something that like really relates to me. Right. I'm like getting old. I'm like, okay, I need to know. Okay, I'm not. I'm not. I'm not. I'm not. I'm not. I'm not that old, but like I'm suddenly to think about major financial decisions.

Okay, sure. I'm like, can I afford a house or am I going to be eternally renting for the rest of my life? Stuff like that. So I'm like, okay, what do I do with my finances? I don't even know the first place to consider or look into. And this demo basically says that it took 20 minutes for this agent to run a task to figure out what a healthy retirement plan might look like or financial plan might look like for this particular user that it is.

is demonstrating this example. And what I found really interesting here, and I'm highlighting it on this tweet, is within 20 minutes, it found local tax laws in Vancouver, which is where this guy must be based. It analyzed average monthly spend rates. It calculated savings needed to retire at 30 years old. It researched optimal investment allocations. It found taxed optimized strategies. It built multiple retirement scenarios. And then it created a downloadable presentation with all the results, Josh. This would have cost me $5,000 from a financial advisor and taken weeks. But here I have chat GPT doing this all for whatever, $50 or $100 subscription. And I'm going to bring up the video here where we basically see it, kind of go through its thought process, and then create a final

finished deck, which you see on your screens right now. It looks pretty rudimentary, doesn't it, Josh? Let's be honest. It looks kind of like a... I wouldn't give it a passing grade in terms of style, but perhaps in terms of the actual content. Very impressive. Very impressive. Very impressive. And this third example that we have is pretty funny because just before we started recording, EJ's and I were going through what we were going to order for lunch. And it was this kind of intensive process. We were reading off the menu items and choosing what we liked. And it took a few minutes.

And this one is a little bit different. This one, the prompt was to order a roast dinner, which I think is very fitting for you to describe EJS because I'm not quite sure what a roast dinner is. This seems like a British thing. It's a Sunday meal, Josh, where the family comes together. It's either roast chicken, roast slam, or roast beef. and it is a lengthy process, all right? Trust me, I've watched my mom do it a million times. Well, according to this example,

it does not have to be as lengthy as you imagine it used to be. What I'm seeing this agent do is it's actually, it's going through a grocery list, it's choosing the chicken, it's scheduling a delivery time, and it's giving you the prompts in which you would need to just finish filling it out. So it's like, hey, what are the login details?

What are the credit card information? What time do you want it delivered? But it's doing everything, and it's amazing. You're watching it click the browser. It is going through the thinking. it is selecting all of the things that you want. Really impressive stuff. So in this case, that solves you your roast dinner problem

where you don't have to worry about it anymore. You set this up once. It knows your preferences. And then you just type in the prompt when you're ready to go and it will place another order. You know what's really impressive about this, Josh? I've been on this website, this exact website, before I actually Tesco when I was a university kid.

And I remember doing bulk orders because I had like no money back then. So I was like, any kind of like bundle deal. I'll order in bulk and then like store it in the freezer or whatever. There are so many products. It's like going to Costco or whatever. It's Costco here, right? Where it's like a bulk ordering thing.

And you know, you just spend millions of hours just staring in the hours being like, I don't know what I need. Do I need this many toilet roll? I have no idea. This just like abstracts all of that away from you. And I just know that there are a ton of people where this would be kind of like super useful for, right? The other thing was that kind of like popular. into my mind is, like, how relevant does this make, like, supermarket websites these days, right?

Or just, like, general retailers in general. Like, do the Amazon's and Tesco's, in this example, of the world, need to now try and cater to these agents? Josh? Like, how are you thinking about that? Yeah, I think the answer is yes. I think directionally, the answer is yes for things much further than just grocery stores, where if you're building a website, you kind of want a version of your website to be readable by AIs. A lot of websites are visually complex. They have a lot of visual. They're meant to be aesthetically pleasing,

but they're not meant to serve data in the most optimized way. So what I've been seeing with a lot of developers they've started doing is they've actually created dot MD file, which is a markdown file that just has the raw text data of the website. No way. No images.

No images at all. Just the raw text data. Because when an AI is scraping a website, all it wants is the tokens. It doesn't care for the visuals. I can't see in most cases. So the most optimal way of serving these models is just by allowing the robot to come and read your website as it wants to. And I think a similar thing is probably going to happen with delivery, with groceries, with everything, is just kind of creating a dual experience, this split experience.

There's an experience you build for the human, and then there's an experience you build for the AI model. And the AI model is going to be really boring and plain. It's just going to be a massive block of text that has no spacing, no prompts. It's just text. But that's really all the AI model wants. It ingests the tokens, it sorts them, and it pops new ones out. And I think this is probably a trend we're going to be seeing with a lot of things, is the user interface, the visual element of it is going to matter less and less.

And I can't talk about this without thinking about an episode that we're going to be filming soon, which is about browsers and AI browsers in general, and how we kind of interact with these things, because it's becoming more and more clear that the future of engaging with the internet will be requesting something from it and expecting to be delivered that thing without needing to search and go through all the tedious efforts of actively going to the website searching through the million different skews that they have on Tesco choosing what you want. You could just say, hey, I have this grocery list or hey, I want to make this for dinner. Just go find this stuff for me and just make sure it shows about my doorstep by 6 p.m.

And that is a really cool trend that I think we are starting to see and we're going to see more as more tools like this get rolled out. I can't help but think that this is going to completely disrupt the advertising industry, Josh. Isn't that like the main way that all these internet companies make all their money? YouTube, Google on search ad revenue, meta, on advertising on all their news feeds and all that kind of stuff? It's advertisers, right? And now I'm trying to think, how do you prove that your advert has made an impression on an artificial intelligent agent? Right?

To your point, you go from making flashy ads, which were designed for human eyes to see and quark at, to agents that are just reading a bunch of dot MD text files, as you put it, right? Which is kind of like insane to think about. The other thing that I'm thinking, the other thing that I'm thinking about is, I think services are going to look very different now, Josh, right? It's going to become less human-like.

So we're not naturally going to go on websites. And I kind of have a comment on your browser, a topic that you mentioned. I kind of think of it as like an intermediary step, right? Almost like it's kind of like trying to bridge the gap between humans and AI. Oh, we're used to browsers, scrolling browsers. So, you know, let's add a few AI features to these browsers. And I'm really excited to do that episode. But then ultimately we're going to end up in this world where it's just advertisers,

trying to pitch agents, maybe your own personal agent, to sell something to you. And I wonder whether it skews incentives in different ways, right? Maybe you can pay the agent, shill a certain product to the individual. I wonder how old of that ends up mapping out. It's a weird thing to think about. I would actually love to have a specialist on advertising on the show to talk about this. That just knows more about it than we do. Because the model is already starting to break.

We've seen it a lot with ad blockers now being built into browsers. A lot of companies are kind of optimizing. out of displaying ads. So on YouTube, a lot of people will just pay for their premium service to not even have to deal with it. So there's this trend towards avoiding ads. Now that these agents are browsing the website, do they insert the ads into the markdown files of the website and hope that they get served? Is the agent going to be able to filter out from there? Because that's this thing called injection poisoning, right? Where you could kind of inject these things into the words that it's reading to give it commands. And that was actually part of the

presentation that Open A had was about the security and defense. defensiveness of these agents. Because when you have an agent that is equipped with your credit card information, all of your personal information, everything about you, and it has read and write ability on the open internet, well, you kind of want to make sure the agent is not giving this stuff up and telling a malicious actor what your address is, what your credit card information is, what all your preferences are. So there's the second side of this conversation, which is the security of allowing an agent to go out into the wild with your entire profile. in its brain, in the hopes that some malicious actor can't prompt inject and create this malicious

intention that serves up the information to a website that they're just, that's not real. There's this whole defensible side of this argument to be had also. It seems like they're doing a lot of work to prevent this, but I guess we'll see as time goes on how it actually plays out. Okay, so so far we've spoken about retail consumption of these agents, Josh, was spoken about the security. One thing that we haven't covered is kind of like the workforce and the enterprise world, right? Presumably, right, if you can connect an agent to, you know, very neatly use your Google Drive, your Slack, your LinkedIn, your Salesforce CRR. What are you doing from 9 to 5 every day, right? You serve as a conductor, I guess. Yeah, I pulled up this, yeah, I pulled up this demo here,

Josh, I remember you can see it. But it's this demo. that shows someone using the GBT agent to basically make personalized invite connections to a bunch of different people on LinkedIn. And so it kind of sets the scenario with the prompt saying, like, hey, like, I want you to basically reach out to individuals that have this kind of a background, that have held this kind of role, that has this many years of experience. And I want you to pitch them on A, B, and C, right? And like, throughout the demo, it kind of like goes through, personalizes each request to

each individual person. And the first thought that I had in this, Josh, was if you're on the sales team of some company, your job just got automated, dude. Because a lot of the time you spent, if you're a recruitment person, work for a recruitment firm and your headhuntick someone, 90% of that time is spent on LinkedIn. I get a ton of these requests day in, day out, right? And so I'm thinking, well, now your job can be completely automated and it could probably do a much better job and reach thousands of people that you, you know, reach over months within an hour, within 20 minutes, within however long this video is, which is two minutes and 36 seconds. So I think that we're going to start to see a lot of these agents in work environments. I don't know if we're going to have as much insight into that.

It's going to be, you know, at the discretion of all these different companies as to what they want to reveal to us. But that's something that we should keep our eyes on, Josh. And, you know, I would love to get someone on this show that knows about how it's transforming their own company, right, at a massive scale. Yeah, it's the, so there's two elements of this agent, right? It's like the personal unlock that we talked about at the top of the show, where now it's this super assistant that you have to take care of all of your irons that you don't want to run, all the things that are annoying. And then you have the professional version of it where it can go and it can make sales funnels and sales calls and it could do a lot of the professional things. It can query data. It can create nice spreadsheets. It can create Google slide presentations. And granted they don't look that pretty, but I mean, directionally, they're heading there and they will eventually get there. So it's this new trend that. is really exciting because it's now here and it's accessible for $20 a month and you can actually go and try it. And you could push those limitations, see where the outer bounds lie and see how it can mostly optimize your life. And then as we go, it'll just unlock more and more and more.

It'll start taking down your credit card. It'll take down your whole preference stack and just become the supercharged agent. So I think that is opening eyes agent feature. It's live. It is out right now. I'm hoping my account gets it soon because I really want to use it. Yeah, dude. You know what? I am officially more optimistic than I was at the start of this episode about this thing. I think you've successfully, or rather, we're both successfully convinced me that this is going to become a much larger thing. And I'm just looking at the end of this demo here, Josh, like this screenshot of like, you know, this virtual browser.

Even though it's so rudimentary and basic, I kind of, I kind of get it. Do you know what I mean? It's like running a simulation on my screen and like having it like just set and forget. and the top left on my screen and just letting it do its thing. Or like if I'm doing work and I want to watch a stream of something, I can have like a YouTube video like there on the corner of my screen. It's kind of similar to that. So even just me saying that, it's behaviorally getting to me.

I'm understanding it subconsciously. And I'm like, yeah, you know what? This is going to be fun. So I'm bullish is what I'm saying. Great. I am bullish too. My hot take before we leave, I'm just going to mic drop this is I think the agent is much more impressive and much more important than the AI power browser.

I think the browser is kind of dumb. The agent hops right over the browser. That's probably worth his host. Dude, that was going to be you. That's going to be my take on the future episode. I was going to show it. We're going to stop. We're going to stop before we spoil it.

The browser episodes coming soon. As soon as we get access to comment, we want to use comment from complexity. We're working on getting access to it. Once we do, we'll have a fully comprehensive episode covering AI browsers. But for now, that was Open AI's agent feature. I would encourage you if you have a plus plan, which is the $20 a month plan. Go check it out. Give it a try.

See where it can automate parts of your life. See where it improves it. See where it detracts from it. I don't know. I'd love to hear feedback. I'm sure both EJAS and I would love to hear more examples of how you're using it. Because at the end of the day, the only real constraint on this is the creativity that you use when you're using it. So I hope you enjoyed it.

That has been another episode. We will be back again soon. Talking about more AI news in the future. I'll talk to you guys soon. See you guys.