Episode 142: ChatGPT Work and Astra: From Chatting to AI Agents

A year ago, I remember thinking about the concept of, like, oh my God, they're going to make an app that's like an agent working on your computer. It is that. It really is. Like, I think a lot of people have missed the boat on this: we have reached agent status.

Hi, my name is Demetri Panici, and I'm a content creator, agency owner, and AI enthusiast. You're listening to the AI Agents Podcast, brought to you by Jotform and featuring our very own CEO and founder, Aytekin Tank. This is the show where artificial intelligence meets innovation, productivity, and the tools shaping the future of work. Enjoy the show.

Hello, and welcome back to another episode of the AI Agents Podcast. In this episode, we're going to dive into ChatGPT Work, as well as the amazing new things happening in ChatGPT. It's an incredible time right now, and we're excited to bring you all the cool new things ChatGPT has done to take the reins against Claude in a lot of ways.

Yeah. Recently, ChatGPT has been doing a lot, especially with Astra. The models are also getting better, including for coding. I was using Claude before, but then I moved to Codex, and it's just faster. This race always has a different winner from time to time. I feel like OpenAI is a little bit ahead right now.

Yeah, agreed. We're going to start by explaining what ChatGPT Work is and pairing that with what Astra is doing because it's obviously the best computer-use model right now, and probably the best coding model as well. I use it a lot for my own coding needs. When you have the Max plan on both, you want to get the most out of the $200, so I definitely test them both out. To kick things off, I'm going to show you the website and where to download it. ChatGPT.com/work is the main place to check it out. I'll have this playing in the background for a couple of seconds. Long story short, ChatGPT does a couple of different things: it chats, it works, and it codes. The confusing thing for a while was that Codex and ChatGPT were two different applications. But about six or eight weeks ago, I want to say, Aytekin, they merged them into the same application. Now they're heavily marketing and showcasing how this is Work, this is Chat, and Codex is in the Work tab. If you ask it to code, it just becomes Codex. ChatGPT.com/work is this.

I remember people were hesitant to use ChatGPT like Codex. They'd say, “I'm not a coder.” They were afraid to look at the code, and the same thing happened with Claude Code. Even though you were just vibe coding and didn't actually write the code, people were afraid to touch it. But with Claude Work and ChatGPT Work, you can give it to anyone, and they don't actually see the code. You're building products with them, doing projects and project planning, and whatever work you do, without seeing the software you're building. That's the amazing part, and it doesn't scare people.

Yeah, agreed. As a starting point, we can showcase this graphic I really like. The site shows some things, but this graphic compares the product across the three options. Let me pull it up real quick. Change tabs. If only they could make an AI for sharing tabs easier. As you can see here, there are three different reasons you'd use it: just answer and explain for GPT. That's your basic “let me ask questions,” interacting with it in chat. Work is actually getting work done. This includes things like research, but it also takes action. It's like controlling your browser. If you say, “Hey, I want to sign up for this new account somewhere,” it'll find the website and start the signup process. Then you just have to authenticate if it's Google. If you say, “I want to set up this thing,” it'll start clicking around in the browser for you, or even control apps on your computer, which is really cool. You're basically building agents as opposed to chat. With chat, you're asking questions, getting answers, and doing deep research. It's about questions and answers. With ChatGPT Work, whenever you make a request, ChatGPT is creating agents in the back end, and you're not even aware you're creating them. You're not aware you're building code or software, or that you're building agents to do the work for you. You're just asking for the work: “I need this.” In the back end, ChatGPT Work spawns AI agents, and those agents do the work for you. That's the amazing part.

Yeah, and it does things for you that you wish you could have done before. For example, you set up a Google Calendar connection and say, “Hey, I'm really busy this afternoon. Can you clear off my schedule?” It'll remove all the meetings. Automation nerds have known how to do this in more long-form ways that took a long time to set up. Now, as the demos will show, you can make websites from scratch just based on an idea, like wanting to get to Disneyland with your kids. That would have taken a lot of research and presenting a plan to the family. As you'll see in a second, whether it's Disneyland or the other cool demos, it's crazy how quickly you can go from an idea to finding things and putting together a proper showcase.

Yeah, I have three different demos today. First, I want to show how you can do project planning with ChatGPT Work. Whenever you need to do something new, you can create a project, give it a name, and start chatting under that project just like you would with ChatGPT. In this case, I said, “We're going to do a launch party in San Francisco. Let's plan and execute this project using an internal website.” The great thing about ChatGPT Work is that it can create sites, products, and sub-products to build websites for you. It created this website for me, and it's completely private, shared only among employees while we're planning the party. We're deciding on the time, date, and location. We haven't decided on the venue, so I asked it to research and find good venues. It browsed and found potential venues, including how many people they can accommodate and how close they are to the office. This is on the Embarcadero Center. It provides different venues so we can decide which one to pick. It took five or ten minutes. I don't have to browse the web to research all this; it's already here. We can have a meeting to discuss the venues, make a decision, and continue from there. The great thing about project planning using ChatGPT's site feature is that you can always go back and look at it. You can use your mobile browser, and it's available to other people so you can build on it. Usually, when you're doing project planning, you're trying to find a document or a spreadsheet with the venues, or the pictures, and it takes so much work. But if you build a website, you can see the history of how you made decisions, what you looked at, and where the project stands. A manager can check in without asking, “What's going on with the project? Did you decide on the venue or send invitations?” They can see that invitations were sent and how many people have RSVP'd. It's a great way to run projects. As I mentioned, it's creating agents. When I asked it to include Jotform in the website, it created a Jotform brand-assets agent. The venue photos agent was probably discovering venues, downloading pictures, and adding them to the site. You have all these agents, and if you don't like the venues, you can ask it to do new research. There's so much you can do. We're doing a lot of projects using ChatGPT Work because websites are a great way to bring everything together. I'm going to show you another one, the Disneyland trip demo. This one is about going to Disneyland next summer. I have my kids' information: my boys like roller coasters, and my girl is only four years old. Let's make two different plans for the boys and the girl, so my wife and I can split up and do different routes. I asked it to plan this, and again, it built a website and started planning everything. The great thing about a website is that you can share it with others. I can share it with my wife, and she can look at the plan and give feedback: “I don't like this.” In addition to websites, you can ask it to create a presentation. For my girl, I'm planning to show her, “Hey, we're planning to do this.” I created a slideshow for her. I'll show it to my four-year-old and ask, “How do you like these?” I can ask for her feedback and see what she likes. Maybe she likes It's a Small World. That's actually my favorite one; I love it. From the website, you can use all the research and assets and data to create different kinds of assets, like spreadsheets. So, yeah, I'll hand it over to you so you can share.

Yeah, for sure. That's awesome. Helping the family with planning is great. I don't want to call things like that not fun, but practically, it takes a lot of time to plan a trip and showcase what we're going to do. I remember my dad would have a piece of paper with the agenda of things we were going to do on vacation. It's so much cooler to open your phone on the trip and ask, “Oh, yeah, what are we doing today?” The trip plan is right there, so it's really exciting.

I actually did this for our trip to Disney World in Orlando this year. I asked ChatGPT to build a website similar to this one, and it did something amazing that I didn't instruct it to do: it created Google Maps links. On the website, I could click the map on my phone, and it would open Google Maps and show me where to walk. It would say, “Okay, first walk to Hulk.” I don't know Hulk; I'm scared of that, but my boys actually loved it. Then we'd walk to the next thing. It made it so much easier and saved us time. We just followed the path the AI designed for us. Yeah.

Yeah, I'm sure it did. I'm going to pull up my example real quick, the first one I enjoy. AI news is a big deal. For my purposes, there are a lot of different sources I used to go to. This is funny because I remember a couple of years ago, before AI was out, I had this conversation with someone in the no-code space. I said it would be great to have all the different websites and YouTube channels scraped and have your own ongoing news hub for everything coming out. Now you can. I have a ChatGPT Work scheduled chat running that scrapes all these sources: Ruben Hassid, a popular LinkedIn creator and YouTuber talking about AI; Matt Wolfe, the go-to AI news guy on YouTube; AI Advantage; Matthew Berman; and all the official websites. Every six hours, it finds the latest and greatest and gives me the top 30 pieces of information, ranked by importance. When Astra released, it gave me all of this, including the original OpenAI post, the GPT-6 Astra announcement, the GitHub release, and a YouTuber's hands-on demo. You can see the immediate post from ten days ago and all that kind of stuff.

Isn't that the best part? Being able to customize it, right? Whatever you like, you can tell it, “I like these particular YouTubers,” and custom-build things for your own needs instead of trying to use all these different places.

Yeah, and there are a lot of tools out there that have always existed, but the problem is that I like things a certain way, you like things a certain way, and it's hard to make something perfect for your needs without a tool like this. I literally just told ChatGPT, “Make a website for it using ChatGPT Sites.” I'll pull it up on the app if you want to see it. It's pretty easy to make, and we were talking about this before the call. You can even have a custom domain if you want. This is the news dashboard chat. It's funny because you remember the whole OpenClaw thing we had before. Whether it's my MacBook Pro now or the Mac Mini I bought when OpenClaw was becoming a thing, I don't think people realize the Mac Mini craze was a good idea, just pointed in the wrong direction. If you use ChatGPT, which I like a lot more than Claude Code and Co-work, you can have full access and interactivity remotely, whether it's Work or Co-work. On my phone, I can interact with it and ask it to code, do work, or set things up. I had this idea on a walk, and it made this. This is the Sites feature, where it builds and deploys sites. You can put in a custom domain. While walking, you can pull out your phone, open ChatGPT, and ask it to do something for you. That's the amazing point. It's like my own app, and it works on mobile, too.

Yeah, this is the first of many things I've tried to do. I'm building and coding a lot of things right now, but for me, this is the biggest step forward. As the days went on, DeepSeek became a product in there, and I thought I should be able to filter by companies, like OpenAI, Claude, Perplexity, and Google. If I want to see OpenAI news or Claude news, I can filter for that. I'm a content creator, so I've built a bunch of skills for making things like LinkedIn copy and YouTube thumbnails. I even made my whole content system, which uses a bunch of skills to make thumbnails and other things. It starts from this one idea I had on a walk and turns into all that content. I also researched different creator businesses, looking at what they're posting on platforms like Skool and Substack to see if I have ideas, like making a little website. For example, I might make a clone of this so people can access it without the content studio connection if they sign up with their email. It could be a lead funnel to get people to sign up for my newsletter, because this is a great news outlet. It's well thought out, with tabs and toggles for access, and you can mark things as read or unread. This could practically be a public site without much work, except for the scheduled chat I have running. It works because my Mac Mini is always on at home, so I can be out walking and say, “Yeah, make this.” You're basically building software and a website without writing code.

That's the amazing part. You can always improve it and keep using the latest version. When you use it, you give it feedback: “Improve it this way. Let's get rid of this part.” That continual improvement makes it so much better. Once you've polished it, you can share it with the world. I saw the site somewhere, but I couldn't find it again. You can even give the site you built a domain name and share it with others.

Yeah, for sure. Some other things that are really cool: you can make a lot of different things happen in ChatGPT. I've been testing out their new feature, which I think is much better than Claude and something a lot of people are talking about. ChatGPT actually has an image-generation model, which is very valuable. ChatGPT 2.5 image generation is really good. For the longest time, I struggled to get the visuals for thumbnails looking good. I made a basic MCP server by typing, “Hey, this is the YouTube API. Can you access thumbnails and search the internet for these things?” It compiled its own MCP server, which now sits on my Mac Mini. From there, I can say, “There are great YouTubers like Matt Wolfe and Jeff Su whose thumbnail styles I like. I don't want you to copy their exact thumbnails, but can you analyze hundreds of thumbnails and learn from them?” Now I have a whole system and a bunch of skills. What's really cool is that it's all saved on my Claude or ChatGPT, but also on my Google Drive because I have the Google Drive connector. I have a bunch of things saved now, including a suite of symbols and things for LinkedIn. You know how people have those nice carousels with custom icons that seem to make an inconsistent brand? I was able to make my own brand by asking it to analyze how these posts work. I asked it, instead of generating some random logo, to make a skill that looks up the right logo on the internet and figures out which logo belongs in the context of the graphic. These are things I did casually; I wasn't doing a bunch of work. Sure, I'm pretty technical, so I can tell it exactly what to do with more specificity. But you can have ideas like this and make really good-looking graphics. This is a thumbnail that five or ten years ago somebody would have said, “Oh, nice, that probably took a lot of time.” I spent time having it research the right things, and then it worked toward my goal. The most important thing for all of you to learn about ChatGPT Work is that you should type slash goal. Slash goal is one of the best things. Everyone talked about Plan, which is important, but when you have an idea and want an outcome, you should say, “Make me a plan that turns into a goal.” It'll write out a whole spec sheet of what the end result should be. It'll make a document describing the goal and what needs to be implemented, and then keep working until it's done. It might take an hour or two. At the end, you get things like this set of thumbnails that I had it run overnight. Look at this.

That's really good. That's a really good thumbnail. It's just checking back to make sure it's doing the right thing.

So it's constantly checking; it's not just giving up.

Right. Usually what happens is that we're doing so many things, constantly moving from one thing to another. Sometimes I'll start doing something, then go do something else, and then something else, and forget what I was doing in the beginning. Things like that happen to agents, too. Using Goal lets you avoid saying, “Okay, I'm done.” It goes back and checks, “Did I actually do this right?” If it didn't, it continues running. It might take a long time, but at the end, it'll make sure it gives you the right answer. I love Goal as well. Thanks for reminding me about it.

Yeah, that's a great point. I think AI is almost like us with ADHD: we all have moments where we get distracted or forget what we're doing because we're switching around. Goal keeps bringing it back to the implementation plan. If you make a plan at the beginning that says what needs to be done and how to get there, it keeps the work anchored. Claude Code isn't great at implementing what's in plans. It'll do really well on Co-work, but for some reason Claude Code doesn't. Even on Co-work, it'll say, “Oh, yeah, sorry, I kind of missed the point of what you said at the beginning.” That's a problem, because the whole point was to do what I asked. If it's off base, it didn't do what it was supposed to. It just wasted a bunch of tokens. People need to remember these models have context windows, and they keep having to compact the context while working. With Goal, it keeps bringing the model back to home base, the north star of what it's supposed to do. When it compacts, it doesn't lose the requisite context that matters. With Claude, a lot of the time, when it compacts, it misses the point when it starts again. That's a big problem.

Yeah, the great thing about Work is that it has all these different functionalities. I know you have a chart showing the different things it can do. Can you share that?

Yeah, I'll pull this up.

All right, let's get the window.

Okay, this has a myriad of different things it can do. Obviously, we just showed you Sites, which is really cool. I told you about plugins, for example, transferring a bunch of files in Google Drive. You don't have to organize files anymore. You can say, “I just made a bunch of icons for my brand. Can you organize them into categories in my Google Drive?” That's super helpful. If you're struggling to understand something, you can use browser or computer use to organize your files, go into a website, summarize things, or tell you exactly what you need to do to implement DNS for a custom domain. It can schedule your chats, which is super helpful. Scheduled Chat made things like my website. I can show you the organized projects on the side for the work and trip. There was another section, too. I showed you some skills I made, like the skill to make nice LinkedIn diagrams or thumbnails. Basically, and I don't know what you think about this, a year ago I remember thinking, “Oh my God, they're going to make an app that's like an agent working on your computer.” It is that. It really is. I think a lot of people have missed the boat: we have reached agent status. Maybe everyone thought they needed to make an operating system for it to do everything, but it just controls the operating system. It's not that complicated. It can control things. We were sharing all these different agents and other companies building startups, like website builders, browser use, and computer use. ChatGPT Work has all those things.

With plugins and apps, you can use other products and integrate them into your workflow very easily. It's really hard for startups to catch up with AI because OpenAI and Claude are building all these products. If you're a startup, I think the best solution is to focus on something rather than major things like a website builder.

Yeah. And I will say, by the way, I'm sure you feel similarly about this, Aytekin: the world we live in regarding cost is significantly better on a token-per-token basis in ChatGPT than on the $200 plans people end up using, or even the $20 or $30 plans. They're really kicking ass when it comes to how much value you're getting for your money. You can run one prompt on Fable in Claude Code or Co-work, and your weekly use is done. I'm like, “What? I didn't even ask for anything that hard.” ChatGPT is very cost-effective at this point, in my humble opinion.

Absolutely. And it's fast as well. It's just so much faster compared to other products. GPT Work is really fast, too. In the past, it could take half an hour, but now it does it much faster. It's getting faster, so I think that's important. It's cheaper, faster, and much more capable. Many things like agents, skills, blocks, or browser use are available without you having to do anything. You're just asking for an outcome, and it uses these different things and handles them for you. Even schedules: you just say in your prompt, “Also do this every day,” and it sets up the schedule and does it every day.

Yeah, that's it. The fact that I said slash goal might have been a stupid thing to say because you don't even have to do that anymore. As I was explaining it, I realized I stopped typing slash goal. I just told it to make a goal, and it figured things out. Even with Claude, if you tell it to make a loop, sometimes it doesn't realize you mean slash loop. For those who don't know, Loop is a feature that keeps working on a plan at a certain interval, or keeps working on itself in general. There are many times I'll say, “Do a loop,” and it won't loop. Sometimes it does and sometimes it doesn't. That inconsistency is extremely problematic if I'm going to bed and thinking, “Can you keep working on this while I'm asleep?” If I ask ChatGPT to work on it all night, it will. If I say it to Claude but don't type slash loop every 15 minutes until 9 a.m., it won't. I know that might sound stupid, like maybe it should just finish the work rather than be lazy, but little things like that make products like this effective.

Definitely. Little things and quality matter. With browser use, you'd ask it to do something real and it would say, “I cannot use your password.” Obviously, you can't do anything without passwords. You need to be able to use my password. Today, ChatGPT can use your password to log into websites and things like that, so it stops constantly arguing with us. We're not doing anything bad; we're asking it to use a saved password to log into a website and check something. It does that much faster now. In the past, browser use was more like taking a screenshot and reading it, which took a long time and didn't do a good job. Today, it's really using a browser, and it's gotten better and better. Inside ChatGPT Work, it's even better. You can use Atlas, but why use Atlas when you have GPT Work? I want to show one example of using other tools, especially browser use. Share one example. Basically, this is Alpaca Trader. It's a product, a website you can use, kind of like Robinhood or SoFi, where you can trade stocks. You don't usually log in to buy things; they give you an API key, but you can also use their website. The great thing is that you can create paper-trading accounts. I created one with $100,000, and prompted ChatGPT Work to create a dividend stock portfolio with a $100,000 budget. I gave it some instructions, but it didn't follow them; it followed better instructions. It decided to invest about $10,000 in each of these stocks, like Johnson & Johnson and Procter & Gamble. We can see it on this end. This isn't a website it created; it's actually Alpaca Market's website. It went ahead and bought these stocks. Now I have a portfolio in paper trading, so it's not real. But I could open a real account and do real trading if I wanted to. I could even instruct it to do day trading, buy low and sell high. It made these orders and bought these stocks. I can see my account with $100,000. I'm about 30 cents down, but I'm investing for the dividends anyway. I now have these stocks, and you can see this stock, ADP, is going up. If I'm planning to get better at investing, I can use this paper-trading account instead of real money and work with ChatGPT Work. Once I'm good, I can start trading with real money. It's amazing, a great example of making money with AI. You're literally making money with AI, and it's so much faster. I just say something, and it does the research and buys the stocks for me. It's amazing.

Yeah, I agree. It's crazy that we're in this world now. A year ago, I was caught up in the whole “Oh, wow, we have APIs and web search inside these tools.” Now we're getting real computer and browser control. My biggest example, and one that saved me hours and hours, is Riverside. It's a great tool, but it doesn't really allow bulk downloading or moving files from Riverside to Google Drive. I was annoyed by that, so I told GPT Work to continuously download and organize all the files onto the shared Google Drive between me and Jotform. I didn't have to monitor it because I trusted it. If I'd asked people who work for me to do this, it would have taken them a bunch of time, and they might have said, “Sorry, I didn't get all of it.” This ChatGPT thing is doing a better job organizing my files than people were. Thank you, OpenAI. Thank you.

If you think about it, it makes things that weren't possible possible. If something takes a long time, you might not do it. Vacation season is over, and during vacation you take so many photos. When I get back, I'm so busy that I don't have time to look through them, pick photos, make an album, or get them printed. I'm thinking I'll ask AI to review the photos and pick the really good ones. I wonder how good it is. I'll work on that and report back on the AI Agents Podcast to see how it does. A lot of things you may not be able to do may be possible with AI. AI agents are real and here. It's like a keyboard—or you can use your speech. It's so easy to create your own AI agents and get work done today.

Yeah, I agree. All right. With that being said, thank you so much for spending time with us today. We appreciate you. Make sure to hit the like button, subscribe, and check out all the places we're on, whether Apple Podcasts, Spotify, or YouTube. Thanks for watching, and we'll see you. Thank you for watching. Bye-bye.