ChatGPT Agent Mode Explained

Topics Covered in This Episode:

1. ChatGPT Agent Mode Overview & Naming
2. Live ChatGPT Agent Demo Walkthrough
3. ChatGPT Agent Mode Availability & Pricing
4. Operator vs. Deep Research Capabilities Explained
5. ChatGPT Agent Virtual Computer Functions
6. Spreadsheet and PowerPoint Generation in ChatGPT
7. Mini RAG-Ready Agents with Data Connectors
8. ChatGPT Agent Security and Biological Risk Classification

Keywords:

ChatGPT Agent, Agent Mode, OpenAI, virtual computer, Agentic skills, Operator, Deep Research, browsing websites, web research, synthesizing information, Microsoft competitor, PowerPoint creation, Excel spreadsheet creation, terminal access, public API integration, connectors, data analysis, image generation, multi-agent environments, retrieval augmented generation, mini RAG, AI operating system, human-in-the-loop, security concerns, biology classifier, biological weapons classification, chemical weapons classification, o3 model, Google Gemini 2.5 Pro, agentic models, AI workflows, editable slide deck, Microsoft Office alternative, AI-powered presentations, spreadsheet automation, cloud-based agents, AI task automation, calendar integration, Gmail connec

Podcast Transcript

Jordan Wilson [00:00:01]:
OpenAI just released what I think is the best agent out there that's publicly available, at least, and I don't think it's really close, at least not right now. So what is this new ChatGPT agent? How does it work and what do you need to know about it? Well, we're going to be answering all of those questions and a lot more today on Everyday AI. What's going on, y'? All? My name is Jordan Wilson and welcome to Everyday AI. This is your daily live stream, podcast and free daily newsletter helping everyday people like you and me not just learn AI, but how we can leverage all the latest happenings to grow our companies and our career. So if that's what you're trying to do, you're definitely in the right place. So it starts here with the unedited, unscripted live stream and podcast.

Jordan Wilson [00:01:04]:
But to leverage all this information and become the smartest person in AI at your company, you gotta go to our website. That's your cheat code. It's your everyday AI dot com. So once you're there, you can go sign up for the free daily newsletter. We're going to be recapping the highlights of today's podcast, but also you can go listen to like 560 plus episodes back on our website for free, the entire back catalog. So you can go learn from some of the smartest people in the world, sorted by category. No matter what you're trying to find, we have the answers on there. All right, speaking of answers, if you're looking for the AI news, that's going to be in the newsletter as well.

Jordan Wilson [00:01:44]:
But let's get to the answers on what the heck is this new Chad GPT agent, how does it work and what are do you need to know about it? Well, I'm going to tell you five things that you need to know about it. So on today's show, here's what we're going to go over. I'm going to give you an overview of what Agent Mode is inside ChatGPT and exactly what it does. I'm going to do a short live demo that couldn't go wrong, right? So you can see how it works a little bit under the hood, and then I'm going to tell you five important things that you need to know. All right, enough chit chat. Let's just get into it. So, hours ago, OpenAI released ChatGPT agent. Or is it Agent Mode? Well, I'm not sure.

Jordan Wilson [00:02:28]:
We all thought that OpenAI finally won the naming battle by calling something ChatGPT agent, but then they also refer to it as Agent mode as well. Well, but at least it's not called, you know, operator V2 or some other random name. So whether you're calling this ChatGPT Agent or Agent mode, it is technically the same thing. So let's first go over a brief description of this from ChatGPT, and then we're going to start with the live demo because it could take a little while. So this is from OpenAI's press release or their blog release, should I say saying introducing ChatGPT Agent, Bridging research and action. ChatGPT now thinks and acts, proactively chooses from a toolbox of agentic skills to complete tasks for you using its own computer. All right, let's try it, shall we? And hey, FYI, I should talk first about availability, because that was just updated, like, hours ago as well. So right now it is only available if you have the $200 a month ChatGPT Pro plan, which I have for other users plus teams.

Jordan Wilson [00:03:41]:
So if you're on the other paid plans now, it's supposed to start rolling out Monday, you know, that could get pushed back again. You know, it's usually one of those things that's like, hey, you know, based on crazy demand, you know, we can't start delivering this yet. So it's already happened once. So, you know, Monday maybe, but it could be even later than that. All right, so, yeah, you do have to be on the ChatGPT Pro plan right now, but probably within a couple of days, it will be rolling out to the other paid plans. All right, so now for the podcast audience, I am sharing my screen. And for people that ask, you can always go watch the videos both on our website. So if you sign up for the free daily newsletter, you can just click it and then go watch the video on our website.

Jordan Wilson [00:04:30]:
Or we do have that YouTube channel thing as well, so just go find everyday AI. So, yeah, a lot of people are like, you know, this podcast stinks. There's no video. There's. There's video, y'. All. So it's. It's just on our YouTube channel or on our website.

Jordan Wilson [00:04:43]:
All right, so now I am in ChatGPT. So the cool thing is this is no longer outside of ChatGPT. It is inside, which is pretty big because Operator previously, which was OpenAI's most agentic offering, it didn't work inside of ChatGPT, it was actually operator.chatGPT.com and that was actually a huge disadvantage manage. And I think one of the reasons that a lot of people, including myself, didn't use it a lot, I maybe used Operator a couple times a week, right. Where a lot of things, I'm using them hourly. Right. And I'll probably end up using my, my ex. Definitely I'm going to be using my whole quota for agents.

Jordan Wilson [00:05:28]:
So it's chatGPT.com uh, and there's going to be a new selection. So uh, if you go to the Tools, so it's, it's not in the normal drop down, that's where I think a lot of people are going to look for. So it's in the tools. So under Tools you'll see a new Agent mode and then it's going to also say how many that you have left and then when that resets. So it's 400 for pro users and then I believe 40 for teams and plus users. So I'm going to click Agent mode. The cool thing that I like is when you enable Agent mode, there's some suggested kind of prompts or things that you can run. Reports, actions, spreadsheets, presentations, et cetera.

Jordan Wilson [00:06:09]:
All right, so I'm going to go ahead and click Go on this one. So all this prompt was fairly simple and we're going to check in on it in a little bit. I said go find the five latest episodes of the Everyday AI podcast by Jordan Wilson. Then create a spreadsheet that has the episode number, title, URL, date, episode category in two sentence description. With that information, please also create a PowerPoint slide deck. So I'm doing something technically a, you know, depending on how you look at it, a four, a four step process. So number one, it's targeted web research. Number two, it's some data entry.

Jordan Wilson [00:06:48]:
Well actually synthesizing then data entry. Then it's creating a spreadsheet and then creating a presentation. All right, so it is going live. There's actually some cool stuff going on on screen that we'll talk about at the end of the show. It's like, it's like a cook show, you know, I gotta put something in the oven, you know, and then check back in on it at the end. All right, so let's talk a little bit more about what this ChatGPT agent actually is, what it does. So essentially it combines Operator, which was ChatGPT's first agent, and then Deep Research, which was technically their second agent, and some new capabilities. So if you don't follow along, don't worry, I'll explain what those are.

Jordan Wilson [00:07:34]:
So Operator was essentially or is a tool that could browse websites, okay? And it does it in a different way than Deep Research. So I'll kind of tell you what each one does and the differences, the pros and the cons. So it actually might make sense to describe Deep Research first, actually. So ChatGPT's deep research, it's really, really good. I've said this all along that in Google's deep research are 1A and 1B. There's like 20 companies now with these Deep Researches, and they're great. So it's a great way to have way fewer hallucinations than a base model, essentially. ChatGPT's deep research, agentically, because it uses the o3 model, it will, you know, you will ask it a question, it will ask you a couple clarifying questions, and it'll go out and research for anywhere from, you know, usually about four to eight minutes.

Jordan Wilson [00:08:26]:
And it'll go to a couple dozen sites, sometimes not quite as many as Google's Deep Research, but it does a really good job. And then as it's researching, it decides it might change path, it might go, you know, run some Python code if it needs to, etc. Right? So that's Deep Research, but it does this very quickly. So it's good for looking at, kind of scanning the text of pages. Operator, on the other hand, is good for looking at pages, right? So understanding the visuals on a page. Whereas Deep Research, it's not really like that. Think of that as, you know, it's like a book scanner, right? Where, you know, operating operator, it literally is able to see everything, see and process and understand, because it uses computer vision, whereas Deep Research more looks at the embedded text. So now Agent, ChatGPT Agent combines both of those.

Jordan Wilson [00:09:24]:
So it does use when it needs to, it uses more of the Deep Research, where it goes to a lot of different websites, but when it needs to, it can also look at something and understand something on a website visually. And this all happens in a virtual environment. So essentially you are using it in the browser, right? But ChatGPT Agent has its own virtual computer, which is really cool. And so essentially there's some new tools and new capabilities which we're going to talk about and explain a little bit more, but to just give them to you, it can create PowerPoints, which is big. It can create spreadsheets, Microsoft Excel spreadsheets, that's big. It can also run code in the terminal on its own computer. That's big as well, because then it can tap into essentially any third party public API. That's also where this thing can get a little dangerous as well.

Jordan Wilson [00:10:25]:
All right, so let's read a little bit more about. And this is from OpenAI's website, so I'm going to go over it kind of quick, but I wanted to hit all the details and you know, they obviously worded it very well, so I wanted to read through this. So it says ChatGPT can now do work for you using its own computer, handling complex tasks from start to finish. You can now ask ChatGPT to handle requests like look at my calendar and brief me on upcoming client meetings based on recent news. Plan and buy Ingredients to make Japanese breakfast for four and analog, three competitors, and create a slide deck. ChatGPT will intelligently navigate websites, filter results, prompt you to log in securely when needed, run code, conduct analysis, and even deliver editable slideshows and spreadsheets that summarize its findings. At the core of the new capability is a unified agentic system. It brings together three strengths of earlier breakthroughs.

Jordan Wilson [00:11:23]:
Operators, ability to interact with websites, deep researches, skill in synthesizing information, and ChatGPT's intelligence and conversational fluency. Then it says ChatGPT carries out these tasks using its own virtual computer, fluidly shifting between reasoning and action to handle complex workflows from start to finish, all based on your instructions. Most importantly, you're always in control. ChatGPT requests permission before taking actions of consequences. Sorry, of consequence. And you easily interrupt, take over the browser or stop tasks at any point. And then it says, you know, in terms of availability, starting with the Pro, Yeah, yeah, yeah, etc. Etc.

Jordan Wilson [00:12:11]:
Sorry, y'. All. I just realized that I'm actually still kind of sick, so I'm a little nasally and all this, but AI doesn't stop, so I can't stop either. So. So let's get to the five things that you need to know. Some of them I just hinted at or alluded to while going over that top blog post release from OpenAI. But first, number five, this is kind of a shot at Microsoft. Yes, I know.

Jordan Wilson [00:12:44]:
Microsoft is reportedly a 49 equity owner in OpenAI, and they've invested reportedly $14 billion. So why would OpenAI kind of tread on their ground, right, and enable people to create spreadsheets and PowerPoints? Good question. You know, there's been a lot of reporting over the last year or so that the relationship between OpenAI and Microsoft has become a little afraid. You Know, they haven't really talked too much publicly on that, but there's been plenty of reporting, you know, trying to renegotiate their contracts. There's been some reported feuds and blow ups and all of these things. But it, I mean that's pretty big. And you know, actually, and we did this in yesterday's, yesterday's show, the first reporting on this was specifically that this was a Microsoft competitor, that this agent was a Microsoft competitor, that, you know, oh, now all of a sudden people aren't going to necessarily have to have a power or, sorry, have to have a Microsoft Office license. I don't think that's the case.

Jordan Wilson [00:14:01]:
Right. I think that's going to impact a very, very small percentage ultimately of Microsoft users. Right. I don't think there's anyone out there that, you know, they're or, sorry, I don't think there's millions of people out there that now are going to be like, okay, I can cancel my Microsoft Office subscription because I can build PowerPoints inside, you know, ChatGPT agent. I don't think it's going to be like that. But for sure this is starting to really tiptoe on Microsoft's ground of being the Office leader, right. And just also talking about moving work from the desktop to the AI operating system. It's something I've been talking about for a very long time and people, I think for the first year I was talking about AI operating system, people didn't understand why.

Jordan Wilson [00:14:53]:
This is why. Right? This is why you have to start bringing all of your processes into an AI operating system. Because pretty soon, I mean, these are the, the quintessential PowerPoint and Excel. Those are the quintessential programs for most knowledge workers. Right. And now that you have Those inside of ChatGPT, think of how much of a leg up that you have, right? Once you, once everyone gets access to ChatGPT agent, if you've already been moving your processes inside of ChatGPT, you're not starting from scratch. But regardless, I do think this is a little shot at Microsoft, but I think some of the reporting on that part has been overblown. Number four thing that you need to know, agents have most normal ChatGPT features, which is pretty cool, right? So yes, it marries Operator and also Deep Research, but then also some of the base modes and features of ChatGPT obviously the model itself, right? So you get the, the smart reasoning and the fast speed and being able to flip flop between the two.

Jordan Wilson [00:16:02]:
But also so it can use image gen. So ChatGPT's ImageGen, which is an extremely popular AI image generating program. So that's huge, right? Especially when it comes to the slide deck. I, I only did one test on the PowerPoint one. So far the visuals weren't great, but I also didn't necessarily tell it like hey, you know, make sure to, you know, have a clear visual on each page, right? And use image gen to do blah blah blah, right. If I worked with it a little more, I'm pretty sure I could get some very impressive visuals in a presentation. And also connectors. That is huge.

Jordan Wilson [00:16:42]:
So that brings us to number three, the thing that you need to know essentially now this is a mini rag ready agent and the fact that we essentially have gone from in the course of like three years tiptoeing on the surface for large language models to now anyone with a paid ChatGPT account or at least starting Monday, right? Or Monday or shortly thereafter, you essentially have rag retrieval, augmented generation, right? So many companies spent multi millions of dollars essentially trying to bring in all of their data first, right? So hey, before you know, your query before ChatGPT or a large language model goes looks at its internal TR first, it's essentially going to look at these vector databases and you know, essentially see your data first and then only pass or only retrieve information in its internal database that it doesn't have in your data, right? Companies spent a lot of money, a lot of time figuring that out. Now it's here in a couple of clicks of a button and it's available agentically. This is huge. So what do I mean by it's mini rag ready? Well, you have the virtual computer tools terminal use all that stuff. But here's the thing, you can toggle off web search, okay? And then you can use ChatGPT's connectors. So essentially you can restrict, which, you know, people might not understand why you would want to restrict an agent, but you can have it only look at your data, right? So canva, Gmail, Google Calendar, Google Drive, Outlook, you know, your Outlook calendar, your Outlook email, and there's other connectors in there as well, that is enormously helpful, right? One of the biggest problems, well there's a lot of problems with agents, especially when we talk about multi agent environments. But one of the biggest problems is well, when a single human is prompting a large language model, human gives one response or sorry, one query, large language model gives one response, an agent. It's a little more difficult, right? There's a lot of things going on under the hood.

Jordan Wilson [00:19:19]:
Most of these agentic queries Take it said between 15 to 30 minutes. Right. So think it's fairly easy to check the truthfulness or accuracy when a single human is using, you know, ChatGPT in a large language model, Chat. When we get to Agentic, it's much trickier. That's why it is important to know that you have these capabilities to actually restrict the web and be like, no, agent only look at my stuff. Right? You still, obviously, you need that expertise driving the loop. Human in the loop. Throw that out.

Jordan Wilson [00:19:58]:
Human in the loop. FYI, so 20, 23, it's dangerous. So number two, same tool set a human has and this part is big. Okay. Whereas before, yeah, you had Operator, you had Deep Research. You have all these different modes. You have ChatGPT, but you know, operator and ChatGPT aren't kind of connected and that Deep research is inside ChatGPT, but it's kind of on its own island now. They're all, not only are they all now together and can share information and share context, but they have a virtual computer.

Jordan Wilson [00:20:40]:
Yeah, that's big, but also scary. And I'll explain that now for, for, for number one. But essentially this is now in, in my very brief experience and reading about it, right. It feels more like for the first time, even though ChatGPT has, or sorry, OpenAI called deep research their agents, they called Operator and Agent. I don't know. Yeah. Was I impressed at the time? Sure. You know, looking back with context, I don't know, those to me don't seem super age.

Jordan Wilson [00:21:20]:
Like, they don't seem like agents, they seem like agentic. Right. Big difference. This obviously feels like an agent, right. Where a lot of times, you know, people are talking about, oh, this agent, agent, agent. And it's like, no, 90, I think it was a Gardner study found that 99, more than 99% of, you know, solutions out there that market themselves as agents aren't agents. They're AI assisted workflows or agentic models. Right.

Jordan Wilson [00:21:55]:
This is an agent. It is literally because it has its own computer and it is making decisions on its own. That's agency, right? You're not restricting it. It literally has its own computer. It can create and save files, it can run commands in the terminal and access third party APIs on its own. Okay, which brings us to number one, the biological weapons classification. So this is something that OpenAI does with all models. They classify it essentially, like, how dangerous is this model? And this is the first time that OpenAI designated a certain model or mode as a high capability in biological and chemical weapon.

Jordan Wilson [00:22:48]:
Domains with every prompt scanned by a biology classifier suggesting genuine dual use concerns. Yeah. To OpenAI's credit in their live stream, when they announced this yesterday, they said it can be dangerous, right. If left unchecked and bad things may and might actually happen. Right. And I think that's important to think about. It's. It's much different when you know a model, right, the o3 model, it didn't get the, the high capability for the biological weapons classification.

Jordan Wilson [00:23:31]:
An agent is scary in a good way, in a bad way, in a what does my job mean? Kind of way. But regardless, bad things are going to happen because now kind of the, the confine of, of, of how, you know, we've been thinking about agentic models and again, ChatGPT's o3 and o3 Pro agentic models for sure. Google's Google Gemini 2.5 Pro agentic models, not agents. When you're working with agents, the capabilities skyrocket, but so does to the cause for concern. All right, so that's a wrap. But I didn't forget. Let's go in and check. All right, we'll see.

Jordan Wilson [00:24:30]:
We'll see what's in the oven. Luckily, I did it once before, so my very first agent prompt, it worked correctly. So let's see if the, the second iteration, if it did it correctly or not. All right, so let's go ahead and take a look here. All right, so. All right, sharing the screen. There we go. So again, my prompt was go find the five latest episodes of the Everyday AI podcast by Jordan Wilson.

Jordan Wilson [00:25:04]:
Then create a spreadsheet that has the episode number, title, URL, date, episode category, and two sentence description. With that information, please also create a PowerPoint slide deck. All right, so let's see. It says opening to confirm request. Okay. Looks like actually this one is still working. Yeah, because I, I'm looking at it. It's not done.

Jordan Wilson [00:25:27]:
So let's see how long it's been working. There's a lot going on under the hood. Okay. It doesn't give the, the amount of time. It might just be finishing up though. Let's see. All right, well, you'll see it's not an error, so. But it's not done either.

Jordan Wilson [00:25:46]:
So it's. It's been going for like 20 minutes. This one's been really working. So let's go ahead and look at the one that did finish on time and by on time earlier. And this one only took 18 minutes, so it correctly went through. And you can also check on the work as well. I'm not going to go through that full thing right now. But you can always click the worked for 18 minutes or worked for whatever when you're done and see exactly what went on.

Jordan Wilson [00:26:19]:
Right. And, and just to recap the one that did finish in time. A lot, a lot going on here. So we have a, a terminal request here. And let's see, you can like replay it from the beginning so it, it's almost like a little video animation. And then you can scroll through it as well. So it shows every single thing that it did. Right.

Jordan Wilson [00:26:43]:
So it first, you know, did a search for everyday AI with my name. It went through, it looked at the 10 episodes and so you can see the actual order it went through. Then it, it, you know, it's going back, it's thinking, it's talking to itself. I need to do this. Then it opened the terminal to start running some code. There we go. Which I believe it used to eventually create the PowerPoint. So let's go ahead.

Jordan Wilson [00:27:12]:
I'm gonna, I'm gonna click. I'm downloaded this CSV. I can't load it up here in the browser, but I'm opening it up on my screen. All right. And everything is perfect on that. So it gave me great job. It did exactly what I wanted. It gave me the episode number, the title, the URL, the date, the episode category, and a two sentence description that it wrote on its own.

Jordan Wilson [00:27:35]:
That's pretty impressive. That's pretty impressive. That's like an intern project, right? Oh, you know, I need to do a better job organizing all our episodes. Intern, go do this. All right, well it did a great job. All right. And then also, let's see here, here's a presentation. I'm actually checking to see if the other one finished.

Jordan Wilson [00:27:52]:
It still didn't finish. Dang. Okay, so this one here's the presentation and it's a PowerPoint. So I can download it, I already checked it, you can open it, edit it, etc. Etc. So the first one, visually not fantastic, but it's usable. Right. The reason I say not fantastic is it tried to cram a bunch of, a bunch of text on a, on a little table here and it's bleeding over.

Jordan Wilson [00:28:18]:
So we have some text bleeding over. But you know, I went in. I think it's actually just a font choice because when I opened it up in PowerPoint it was fine. But you know, we have the COVID slide here. It says latest everyday AI episodes. It has a dual pane on the right hand side, there's a little photo of a microphone and A, you know, AI kind of graphic illustration, a couple pages for the episode overviews. So nothing crazy, but it, it did its job, right? I didn't tell it like, hey, make sure each page has, you know, visual and a headline and a subhead and. But it's really good, right? Also, I love the, the PowerPoint preview inside of Ancient mode.

Jordan Wilson [00:29:05]:
It's super clean. It feels like you're editing in actual, like editing in a, in a, you know, PowerPoint or a slide editing program. It is super slick and it works really, really well. So to recap, what the heck is Agent mode? Well, it's essentially putting together the best of Operator, Deep Research, ChatGPT and some new tools, right? Having use or being able to create PowerPoints, Excel spreadsheets and also having access to a terminal, which opens up a world of possibilities. You saw in the demo. Pretty good, right? First. First attempt. Obviously no one's going to be using that, that PowerPoint, but, you know, with a little bit of work and more description, it would, it would have probably turned out pretty well.

Jordan Wilson [00:29:59]:
And then real quick again, the five things that you need to know. Number one, kind of a shot at Microsoft. Number two, agents now have most normal ChatGPT features. Number three, there's essentially mini rag now ready working with connectors. Number four, it has the same tool sets as human and it feels more like working with a then in large language model. And then last but not least, yeah, that kind of troubling. Biological weapons classification. Yeah, this thing is very powerful, but also kind of dangerous.

Jordan Wilson [00:30:30]:
All right, that's a wrap, y'. All. I hope this was helpful. You know, if, if, if you do want to do a deeper dive on ChatGPT agent, just let me know. Just, I don't know, type agent or something. Or if you're on the podcast, you can comment on the podcast. I can't reply back, which is weird, but I'll see it. You know, so I always say like, oh, if a certain amount of people will, you know, it's always like an arbitrary number.

Jordan Wilson [00:30:56]:
I'm like, oh, if, you know, 10 people say agent, that means that there's enough appetite. So, you know, if, if you want something more in depth, we could do it for one of our, you know, AI at Work Wednesday shows or something like that. But let me know, just, you know, type in agent so I know if there's an appetite or not. I'll probably put something in the newsletter as well. Speaking of that newsletter, I hope, hope this episode was helpful. If so, please go to your everyday AI.com Sign up for the free daily newsletter. We're going to be recapping today's show as well as all the other AI news that you need to be the smartest person in AI at your company. Also, if you're listening on the podcast, appreciate the support.

Jordan Wilson [00:31:29]:
As always, please subscribe to the show. Leave us a rating if you could. It takes two minutes, but it means a lot to me. So thank you for tuning in. Hope to see you back for more Everyday AI. Thanks. And that's a wrap for today's edition of Everyday AI. Thanks for joining us.