Getting Started with Codex by OpenAI: The Future of Business

Before we get started, I want to ask you something. Are you tired of feeling like AI is moving faster than you can keep up? Because I know what that feels like. And I hear it all the time. Marketers keep telling me they're drowning in all the new tools and updates and features and must-try frameworks. And, and they got to do all this while they're also doing their job and trying to run their business. And that's why the AI Business Society was created. It's your personal curation team for AI. We tell you what deserves your attention and what you can safely ignore. And inside you get live trainings from leading AI practitioners and interactive skill building sessions, bite-sized learning libraries when you just have a few minutes, and a community of incredible marketers who are just like you. Hundreds of them are on the inside, and you can join right now and lock in an incredible special sale we've got going on. Visit socialmediaexaminer.com/ai. Give it a shot. You got a 30-day money-back guarantee, so you've got nothing to lose. Go to socialmediaexaminer.com/ai and let's now jump back into today's show.

Welcome to the AI Explored podcast, helping you put AI to work. And now here's your host, Michael Stelzner.

Hello.

Hello.

Hello. Thank you so much for joining me for the AI Explored podcast brought to you by Social Media Examiner. I'm your host, Michael Stelzer, and this is the podcast for marketers, creators, and business owners who want to know how to put AI to work. And before we get to today's episode, even if you're not interested in Codex, you should listen to this episode because so much of what we talk about is applicable to so many other kinds of technology that we've been talking about on this show over the last couple of weeks and months. It's applicable to Cloud Code, it's applicable to a lot of the other tools. So I think you're going to find this very, very important, which is why I spent a few extra minutes on today's episode. Let's transition over to this week's interview with Carl Yeh. Today we'll explore how businesses can get started with Codex by OpenAI without any technical know-how. My special guest is an AI strategist who helps traditional businesses successfully implement AI and train their teams. He's the co-founder of Zero to Sixty, spelled zero to six zero, dot AI, an AI Consultancy. He co-hosts the AI Accelerator for Business show. Carl, yay, welcome back to the show. How you doing today?

Great. Thanks for having me back, Michael.

Super excited to have you back. So we're here to talk about Codex. Codex by OpenAI, the company behind ChatGPT, is effectively a tech stack that allows you to connect all the other things together and accomplish really powerful things. That's kind of how I would probably describe it. And Claude Code is the equivalent of it but it happens to be Biontropic, which is the company behind Claude. So we're going to get into today how you can use Codex and how you can use this in particular in business applications to really allow a business to do things that are beyond their wildest imagination. Now, before we get into that, what are some of the big misconceptions about Codex? Because it sounds a little scary. Yeah, just the name alone, right? So what's the word on the street about this thing?

Well, even the name itself, right? Codex. So the first thing in, like, the first thing when people hear about Codex is, oh, it's only for people who know how to code, right? Um, I think that was the similarity between Claude Code, right? Just like, oh, what am I gonna do with, you know, I don't code, so what about this? And it initially was built like that to essentially to help people who are software engineers, people who know how to code develop applications and so on. But Very similar to what I think people are finding out about Claude Code is that you can do non-technical tasks in Codex, right? And this, that leads into the second misconception. It's like, oh, it's very difficult to use. Not really, 'cause essentially you can use it exactly how you would ChatGPT, right? So if you type, maybe you speak to your ChatGPT, to your Claude, very similar here. But I think the big difference is, you know, like in ChatGPT, like in Gemini or Claude on the web app, you would have conversations, it would respond to you, maybe create documents for you. But where Codex is different is actually it can complete business-type activities, things like reviewing your spreadsheet, making edits to your spreadsheet, creating brand new spreadsheets, deleting a spreadsheet in that same folder that it's connected to. And I know we'll get to that in a little bit, but just a little bit of taste of— think of Codex as the application or doing, and think of ChatGPT more as the conversation-based.

Got it. Okay, so what I'm hearing you say is despite the name Codex, you do not have to be a coder to benefit from using this tool. And that's a big misconception. And in addition, it is not technical. It's actually quite user-friendly is what I heard you say, right? Because you can actually interact with it in the same way you interact with ChatGPT, which I think is absolutely fascinating. Now, I will tell you right now, I've never used Codex, but I am curious when Codex is used well, and I know you've been able to do this for yourself and for clients, what is the upside? What are the benefits? What does it make possible?

Yeah, so one example would be, you know, we've heard about automation tools like Make, Zapier, n8n. One of the things that, for me, I think it replaces a lot of those workflows because you still have to connect to whatever system or tech stack that you have, you have to, but it's very deterministic.

With Codex, Explain deterministic for people that don't know what that means.

So deterministic is more like if you want a specific output or you want something that looks ex— you know, if you've used Zapier, Make, or N8N, it's very like, here's a module connected to this, connect to this, connect to this, connect to this. And when you're building all that out, sometimes it looks like a whole bunch of octopus arms that's coming out of it. What makes Codex a lot easier is you you can just chat with it and ask, hey, this is what I'm trying to do. Let's— let me build the connection. And how can we build that to accomplish this task? And in fact, recently, a major update was what's called the goal feature. So if you just put /goal, and I know this came out for, um, uh, open— I mean, Claude Code as well, like, I think yesterday. And, uh, For those who use personal assistants like an Open Claw or Hermes, they have goal as well. You just need to know this is my goal. And if you give it the correct instructions, you know, really break down what you're trying to do, it will work hours, if not days, to accomplish your goal or until your credits run out or your token, you're rate limited.

Wow. Okay, so let me just make sure I understand you correctly. If you use Codex and you give it a goal— and by the way, you said this is also now pretty much available everywhere else, this concept of a goal— if you give it a goal, it effectively will work with you to achieve that goal, even if you have no idea how to get there. Is that what I'm hearing you say?

It isn't just work with you to get that goal. Like, just like, you know, proper prompting, you want to be very specific in what you want to accomplish. Yeah, but once you've executed it, Codex will essentially work until it reaches that goal. Oh, so you, you're starting to see people 8 hours, 10 hours, 12 hours, 15 hours until it accomplishes the goal you have set for it, which opens up a lot more opportunities, especially in business-type tasks.

Wow, that's really cool. Okay, so what I heard you say is one of the big advantages to Codex is you do not need to use Make, Zapier, or N8N. And for folks that have been listening for a little while, we've had a number of guests on the show talk about all of these different platforms, which effectively are platforms that allow you to do this if this and that kind of, kind of chains of commands. And they're very particular. And I've been told by a lot of people they break if something goes wrong. Right. And it sounds like Codex is smart enough to kind of like not to accomplish the same thing. Is effectively what I'm hearing you say without having the same limitations.

Is that true? Yes. Yes. Now I know later in the, the episode here, I'll go into a very unique example of another release that came out last week too about how, you know, we, we gave it this goal. You cannot give that type of goal to something like a Make, a Zapier, or N8N just because of what entails all the different steps that require to do it. But when we've done it and we've replicated this multiple times, specifically in a client environment, it has accomplished that task over and over and over with no issues and with unique things that we have set for it because of the situation. So we're starting to see, and you know, this isn't just focused on Claw— um, on Codex or Claude Code or Hermes. I think this is broader where you're starting to see the actual uses of agency where an AI agent can actually take agency to accomplish whatever the end task that you give it versus like an NNMake where we were kind of relegated to really defining the step by step by step by step versus now here's steps, but it's gonna continue until it accomplishes the task versus it'll break the first thing it, you know, it's not— it has no connection or something happens to it.

Okay. All right. So what do we need to be thinking about before we start Codex? Because I can assure you, most people listening to this have no clue. They've never used Codex before. So just kind of like walk us through the stuff that we need to wrap our brain around before we get started with Codex.

Yeah, for sure. Now, if we've already talked about understanding that it's a harness, right? And that by default it will have— because it's created by OpenAI, it's going to have GPT, GPT-5.5 as the default. You can change models, but remember, you can also switch. The biggest piece here, what makes Codex amazing, is the ability to connect to your local either computer or in business cases, your local network. So many, especially non-traditional or non-tech companies, a lot of the work is still being done on their local networks or local drives, right? They save, they work in there, right? Sometimes it's in SharePoint, reluctantly, but a lot of times just by virtue of how many years they've been in business, a lot of the work happens there. And with Codex's ability to open specific folders— so when you're starting a new Codex instance, let's say you've downloaded, which it can be both for PC and for Macs, so you've downloaded it, you can connect to really any folder your computer is connected to.

Just real quick, when we say download it, like most people that are using ChatGPT are using the browser, I'm presuming. So, so there's a dedicated app. Is it, is it an app called Codex or is it part of ChatGPT?

This is where it gets a little confusing because if you are paying for ChatGPT, there's a little Codex on the side that you can use, right? Okay, but what, what I'm talking about here is downloading the actual app, whether it's in PC or whether it's in Mac, called Codex. Yes, called Codex. Okay, that— this is the feature you can connect to your local drives.

I see, keep going.

Yeah, so once you connect it to your local drives, it has— in addition to your local drives, it can connect to your other tech stack. So things like your email, uh, so if you're in on the Microsoft environment, um, Outlook, your calendar, your Teams, um, SharePoint folders, if— or OneDrive. If you're in a Google environment, um, Gmail, uh, Google Calendar, uh, Google Drive. And then think about the combination there, both your local drives plus your, um, like even just your email, your calendar opens the door for so many different, um, work possibilities. And then Now, because you're connecting to your local drives, I'm sure some of your IT-minded listeners will probably think that's a big security issue too. And you have to be very careful because you're giving Codex and any other agents that you create out of Codex access to that folder. And when you connect it to that folder, it technically— you can ask it to do anything in that folder and any subfolders that it contains.

So is the warning here to be careful about not giving it, um, access to all of your working files, but to maybe like create a special folder and drag some files into there so that it doesn't accidentally delete everything or change things without your awareness? Is that kind of what I'm—

yeah, yes. And if you're starting out with Codex, what I would recommend, like you said, create a dedicated folder, um, that you can connect Codex because once you have it, you can theoretically connect it to your root folder, which then gives Codex access to your entire network. So you have to be very careful in how you connect it. And I think this is where, you know, you probably want to talk to your IT or whoever's manning your entire network to, to tell them what you're doing. Because remember, when you start doing work in your folders, on their end, what they're seeing is you doing the work, not your agent. It's just you. So if you're deploying all these different tools and agents in multiple folders, it's going to look like as if you're doing all the work. So very important to have that conversation. Make sure people know what you're doing.

Yeah, we're going to get to some of the security stuff a little later in the interview. I want to talk a little bit about the interface. For Codex. I know that ChatGPT uses quote unquote the terminal, which is really scary for a lot of people. Is it similar with Codex, or what does that interface look like, that user experience?

So when you download the Codex app, it looks very similar to ChatGPT. A couple different, um, titles like plugins and automations, but when you open a specific folder, very similar to how you open a new chat thread in a ChatGPT, it'll show a new chat thread and the work that you're doing in that thread. So, you know, if you have a folder called test codex, you're gonna see test codex and all the different chat threads you have under text codex.

Got it. Okay, let's talk about some examples just to help people understand what is possible here. I know there's a couple things that we talked about when we were prepping for this. I would love you to share these examples.

Yeah. So one of the examples we, what I have is, um, so we're working with a construction company and previously they do what's called invoice reconciliation. So invoice reconciliation is when you take a vendor invoice, in this case, hundreds of them per month, they get them from various vendors and they have their internal spreadsheet that someone manually has to look invoice to spreadsheet to make sure the numbers are accurate, whether it's paid or not paid, there's no duplicate payment, and to make sure that there's no outstanding balances or any discrepancies. That's the big piece. Now, with this company, that took about 2 to 3 full days every month to do that invoice reconciliation. Now, about a year and a half ago, the initial AI solution would be, because they're on ChatGPT for Business, they would upload the invoice into ChatGPT, then they would upload the spreadsheet, and then have ChatGPT essentially do the reconciliation for them. And that knocked off the 2 to 3 days to 2 to 3 hours a month because they set up the double-check, spot-check, and so on. Now what we can do with Codex, and we've implemented this already, there's a folder. It's pretty much a weekly invoice folder, and the team just loads all the vendor invoices into the specific folder that's named after the vendor, whether it's like Rona, Home Depot, whatever the name of the vendor is. They would just drag and drop all the invoices in there, and on Sunday at 10 PM Codex, it runs on an automation, would spin up a sub-agent that attached to what's called an invoice reconciliation skill, 'cause there's a specific way they do invoice reconciliation. It reconciles every single invoice against the spreadsheet. Now what makes this kind of neat is that sometimes there's 5 vendors, sometimes there's 10, sometimes there's 20. But what Codex does, it duplicates each sub-agent based on how many folders there are. So if there's 5, there's gonna be 5 sub-agents. If there's 10, there's 10. If there's 20, there's 20. And they do the reconciliation all at the same time. Mm-hmm. So what you get is each folder gets a reconciliation report. There's a master reconciliation report, and all this happens on Sunday. So all the team has to do is again, drag and drop all the invoices they receive, and on Monday morning they get a fully reconciled report. And again, what used to take 2 to 3 days was then dropped down to 2 to 3 hours, now is 15 to 30 minutes.

Wow, that's really cool. And you also had another example, uh, with the meeting prep. Yes. Tell me a little bit about that example as well. That just might help other people too.

So personally, what I do is, just like everyone else, in the morning you have your emails, you have to take a look at your meetings, then you have to take a look at your notes. So with Codex connected to my Outlook and my calendar, I would get a daily meeting brief. And in that daily meeting brief, it would tell me, hey, any unread emails that I've had, especially because it, once it gets starting to know you, it understands like the specifics of what you're looking for. Then all the meetings you have. Then based on those meetings, any notes you take, you took from the company. So I use Notion, so it's pulling from my Notion too to review any of the notes I took on those particular companies, the meetings I have with the companies they work for. And then also looks into our Slack because all our clients are, has their own channel. So it's pulling from that too. So I get sort of a daily meeting brief plus a meeting prep package, all 6:30 in the morning, and sent to both my Slack— I have a channel dedicated to daily briefing— but also on my iMessage because Codex has a feature called computer use. So it opens my iMessage and puts the daily meeting prep package in there and sends me an iMessage as well.

Hey, before we get back to the show, I want to talk to you about something I've been thinking a lot about lately. You already use AI, and if you're like me, every single day you're using it, and you're not someone who needs to be convinced that it matters. But here's what I've learned from working with hundreds of marketers inside our AI Business Society over the last 2 years. The problem is almost never AI. It's trying to do it all on your own. You're testing tools that you're likely not using again. You're rewriting prompts because they don't sound right. You're probably watching tutorials that give you lots of ideas, but you have no idea what to do with it. And you're working harder with AI, not smarter. And that's exactly why I built the AI Business Society. Every live training is led by an expert practitioner who uses AI in their marketing all the time. And I'm personally there on every single one of these training sessions to make sure that nothing gets too crazy or abstract. You leave with something you can implement that very day. Plus, you're surrounded by a community of marketers who are just like you, and they're all there to support each other, and you get to be part of this awesome community. The investment is $497 for the full year, and if within the first 30 days you decide it's not for you and you join, then just ask for a refund. We're not gonna ask any questions, but at least you gave it a shot. Head over to socialmediaexaminer.com/ai. That's socialmediaexaminer.com/ai. And now let's get back to the show. By the way, folks, this stuff is absolutely fascinating. And the more that I interview folks like Carl for this show, the more I'm beginning to understand that everything we're talking about today is doable with Codex. It's doable with Manus by Meta. It's doable with Claude Code. Like, we're beginning to see the standardization, if you will, of what we're talking about today. And it seems as if, correct me if I'm wrong, Carl, that almost all these tools have the same connectors and all the kinds of skills. And we're opening up this world of possibilities right now, which just seemed nearly impossible a little while ago. And it's kind of mind-blowing. It absolutely is mind-blowing. So I just wanted to throw that out there, folks. Even if you don't use Codex, you're going to learn what from stuff we're talking about today that's going to be applicable to one of these other tech stacks. But we're here to talk about Codex. So I now want to ask this question. So now that we've talked so far about like, okay, it's a harness, right, for ChatGPT, and you interact with it in the same way you interact with ChatGPT, and you can, if you want to, have it use Claude instead of ChatGPT. And you've got these local files that live either on your hard drive or on a network. And you can do cool stuff like we just talked about, invoice reconciliation or these meeting preps. Now what I would love to do is actually get into a little bit more about the actual use of the app itself. So, so far we know that there is an app that you should download. You mentioned earlier that there, that Codex is also built into some other app. So help me distinguish between, is it built into ChatGPT app and it's also a standalone app? Help me distinguish that a little bit here and then we'll get right, right into the desktop app and we'll start talking about skills. And all that fun stuff.

I think the confusion here is there's ChatGPT by itself, right? Most people use it via their browser. Yeah. There's also a ChatGPT app by itself that you can use. And what I, I say is that app is relatively useless because most people use it on their browser. And then you have Codex. So what, what does the purpose of a ChatGPT app actually have? So I think that's the confusion. Now Codex itself can be used similar when people talk about, you know, using Cloud Code in your terminal, it can be used in your terminal too, very similar use. But I think the big differentiator is the app because everything you can do in ChatGPT, you can do in this app and more like what we've just talked about. So I think that's a, that's a big difference right there.

Yeah. And I'm hearing people like Deanne Sanchez talk about how he's not using ChatGPT anymore. He's using, he's using Codex because it's ChatGPT like on steroids. That's kind of how I think I'm hearing him describe it. So let's, let's get into some of the things. Let's talk about skills a little bit because I think we need to wrap our brain around that.

Yeah. So things like skills, prompting, MCP's connectors, that's what's called the scaffolding, or I like to call it the tools that you're giving to your robot, right? Okay. So the robot has an, like, internal tools that it can use that Codex enables it to do. But to make it work with your data and make it work in your business environment, you obviously need— still, prompting's important. Building skills with it. So one of the things, the difference between a prompt and a skill is a prompt is you, you would use a prompt for like a one-time use, right? So like one-off activities. But now because in most businesses everything is repetitive, there's things that you do on a regular basis, weekly, monthly, daily, whatever that is, this is where you can create a skill. And a skill is literally just teaching your AI tool, your harness, your system, whatever that is about your process. And that skill, let's say a simple skill, and let's take a marketing example here, is we all know there's some sort of brand guideline for most businesses. And most companies have a brand guide, right? PDF brand guide, whatever that is. Some people have writing style guides. So, One of the best uses for a skill is to turn your brand guide or writing guide into a skill so that there is a standardized skill that you can provide to your entire organization. So whenever they're creating any documents, anything, they apply that company skill, writing brand. And as once you've, you know, dialed that in and, you know, worked through the approval process, everything they create using AI will be adhered or adherence to and outputted exactly how you want the company brand or writing style to be. And so we've implemented that many times over with many of our clients. That one of— that's one of the easiest way to deploy because how many times have we heard from marketers, you know, people are creating all these things that are not abiding by the company brand. But if you give them a skill and they're using AI, they'll just literally drag and drop that skill, whatever harness they're using, and it'll come out with your company brand.

Love it. And when we were prepping for this, you had a couple of cool skills that were not created by you that you're using. I believe one of them was called Playwright. Yeah. And another one was called ReMotion. I would love you to explain what these skills do and where we can find these kinds of skills also.

So I actually want to add to that, and this is one of the most useful skills or plugin that you'll ever come out of that's even more important than ReMotion or Playwright. It is the Chrome plugin skill that was just announced last week. Okay. So there's two things you have to do. One is you have to go to Chrome, um, and download the Codex extension, right? So go to Chrome extension. Download that. Then you go into Codex and download the Chrome plugin. It's in the plugin section on the top left there. You'll see it. It's pretty much— they put it up right at the top. Now what this does, it allows Codex to control Chrome.

That's a huge advantage, right? Because almost all of your apps are web apps, right?

Yes, yes. And what makes it even more powerful, most like, you know, when I was talking about Playwright or Browser, when, or even Claude in Chrome, what many of these things do is take screenshots of your browser as it tries to accomplish whatever task you're trying to do. It's a big difference on how Codex actually runs Chrome. And I'll give you an example of that. Many companies, not tech, so traditional, they're using tool legacy software, things like Trimble, Bid2Win, Yardy, for those who understand this, iFrame or iFirm Caseware, things that sometimes even look like SharePoint 2010 interfaces, right? Very traditional old software. What many people do is they go into the software, extract that information into, let's say, Excel, and then work in Excel and then import back into that software or another legacy software. Any non-technical business or non-traditional businesses or traditional— sorry, traditional businesses, they do lots and lots of that. We tested Chrome plugin last Friday on such a system. And so, what a person normally does, like in this case, a construction company, this foreman would have to go into the system, log in, then about 7 clicks in, have to filter to find a specific, in this case, field log. It, you know, every day, you know, the company writes field logs and they have to find it and they have to export it into Excel. So that takes that person about 8 to 9 minutes. Why it takes that long? Because it loads— it's very slow loading. You click one, it slows, and then if you get to the filter part, it's really slow. So what we did is we used Codex— I mean, um, the Chrome plugin in Codex. They've already logged into this system So Codex just opened a brand new tab, and the instructions I gave it were, I want you to create 3 subagents. One agent to go to this particular page, go to this filter, filter it to this field log number, and then extract, hit export to Excel, and save that Excel file into this specific folder. That was Agent 1. Agent 2 is, once that's complete, I want you to add a formula into that Excel. This is the exact workflow this company does. Add a workflow into that Excel because they need to change the number. The third agent goes into that final Excel folder and updates that Excel output based on another Excel that is in the current folder that we want to save it into. So 3 different agents, 3 different tasks. We have run that multiple times and it has completed it. Took about 18 minutes per— 18 to 20 minutes per completion. But I know the owner sat there and saw it and he's like, that has literally changed how we will do our business. Hundreds of hours because our team is so bogged down and going into that system, extracting itself, putting formulas and doing a comparison that you have just shown, like, in that 25 minutes, that agent has done like half our task.

I love it. And obviously you can have these agents working in the middle of the night when you're not even on your—

oh yes, yes, which is really cool.

Okay, so briefly describe Playwright and Remotion so people understand what that is.

Just real briefly. So Playwright is the skill. Yeah, it's a skill. It's— and that can act— essentially what you're trying to do is have Codex go to different websites. And there's a difference. It's called headless mode versus head mode. So headless mode is it's going to these websites without taking, uh, screenshots. It's going through this, and, and it's really helpful if you need to scrape information from websites where it doesn't need to essentially open a browser, open tabs. It doesn't need to do that. It just scrapes that information and, um, gives it back.

And for folks that are curious, it's spelled W-R-I-G-H-T, Playwright. But what I'm hearing you say is the Chrome plugin now kind of makes that potentially obsolete.

Makes it really— now remember that the Chrome plugin is a little slower. It's still quite fast, but if you just need a straight scrape of website, I would still use something like Playwright.

And what is Remotion?

And Remotion is the ability to create animated video just using essentially natural language. So it'll actually create videos for you and, you know, you can upload your files and whatnot to create those animated, especially text-based videos or even vector files. And what's really nice about this is because you can, because Codex has the ability to create images using the new ChatGPT ImageGen 2, so you can combine ImageGen2, create images, and then have ReMotion create the motion, not like essentially animated videos from it. So you've probably seen those non-talking head, just voiceover videos. It can create those type of videos. Very cool. Mind your token limit though. Mind your tokens.

Where do we actually find all these cool skills that are already pre-developed, if you will? Are these from the makers of the software? Is that generally how that works, or is there like a Is there like a directory that is a trusted source inside of Codex that has all these skills? Yes.

Okay. Yes. So there are— there is a directory in Codex. It's under plugins. You can see all the plugins and skills that are there. One thing to note, I know we're getting security quite soon, is be very careful when you start going to different marketplaces when you're downloading these skills because there is— you have no clue what is in these skills or plugins, right?

But you can trust the ones that are in the Codex Marketplace.

I think you can trust, you can definitely trust more of the plugins from the Marketplace on Codex. Yeah. You've probably seen so many people promoting their own marketplace or their own skills. Be very careful about downloading those.

Okay. You talked about computer use earlier. Explain what the heck that is and how that's different than what the Chrome thing is.

Yeah. So computer use on Codex enables Codex to actually open and control applications on your computer. Now, it's only for Mac, but the example I gave was, you know, in my daily brief, it would open my iMessage and text me my meeting package.

Okay. All right.

Yeah.

Well, you also mentioned Excel. So I'd imagine if you had Excel on your desktop, it could open that as well, right?

Or you could open Excel. It can open Word. It can open— now, there's different ways to access it, right? So like, would you want it to open Excel or just go into the folder and work in Excel? Like there's things, there's now many different avenues to complete a task, right? So you could have it open Excel or, you know, go into the folder and start doing the work in there without opening it too. Some of the things that are really neat about it, especially when you get into the coding aspect, is, you know, opening, um, like you can open Chrome. Um, well, actually you wouldn't open Chrome, you would open something like you mentioned Notion. Yeah, you can, you can open your Notion. Here's the, here's the confusing part about that, right? So like in computer use, you could definitely open Notion, but then you can also, because it's connected, you have a connector to Notion, you can just, you know, chat in Codex and then it would automatically go into Notion without opening the application itself. So it starts getting confusing and what you can do open. I had it actually open my Final Cut. And that's when it starts getting a little, I'm pressing too hard because I said, can you try to edit this video?

And it's not quite there yet, huh?

No, not definitely not quite there.

Well, I would imagine Adobe eventually will have some sort of plugin that will allow it to be in Premiere or Apple will come out with some sort of plugin. I mean, this is the future, right? Where all of these major software providers are going to have either plugins or skills or both that will allow all of these AI models to be able to do some powerful things. Now, you hinted earlier about automation, and I really want to get into the automation side of this because— and you've been— maybe you've even been talking about automations a little bit, but talk to me about like automations in general and what we need to know and how all that stuff works with Codex.

It's actually not that hard in Codex. So for example, the, the example I gave you was the invoice reconciliation that happens on Sunday at 10 PM. What you do is, for us, it was, it was an iterative process, right? So you're like, oh, we have to reconcile one invoice first, make sure the reconciliation is good, create the invoice reconciliation skill, make sure the sub-agent uses that. But once we're comfortable with the entire process and we've built it out manually, manually as much as we can, then in Codex, all you need to really say in the same thread would be, so there's two type of automations. I'm gonna talk about the first part. The first part is, It spins up a new thread every single time. So think about a new chat thread. So every single time there is an invoice reconciliation, it would spin up a new chat thread. It would do the reconciliation. You can see it on your left-hand panel, similar like you would in ChatGPT, a new chat and all the reconciliation it did. You know, you can, then you can chat against it. The other automation you can do is what's called the heartbeat. For those who've used OpenClaw, Hermes, you probably know what Heartbeat is. It'll essentially codex your agent, turns on, does the task, and then goes back to sleep.

On a specific interval, right?

A specific interval. What makes this even better is it does it in the same thread, so it doesn't open a new thread, so it keeps the context. An example of what this could be is for me, when I'm really deep into work, I don't really want to be, you know, bothered with trying to look at emails that I missed or Slack messages. So what this does is I turn on this essentially checker every, whether it's an hour, every 2 hours, it would give me a summary of all the unread emails and unread Slack messages and anything that's important. And it's learned this because it's in the same thread. So every couple hours,, it would send me usually through iMessage, hey, here's the, here's the emails you missed, here's the Slack messages you missed, and here's the things that it thinks that, that are important. So that would be example of a heartbeat.

I think it's important for people to wrap their head around this a little bit because if you think about human-to-human work, typically what happens is humans are waiting to be tasked to do things by other humans, right? And that usually could be done in our case through Asana or through an email or through a chat or whatever, right? Like, hey, this is ready, go do your thing. That is effectively what you can do today with a lot of these things. It sounds like Automation A, which is like human A does their thing and then brings in human B and then it does its thing and it's a series of activities. But what I like about this heartbeat concept is it's kind of like the other thing where, you know, you have a weekly report that you have to do and ideally you've got a weekly reminder on your calendar or in Asana that's a repeating task to remind you to do it. You can set up these heartbeats for all sorts of things, right? So, so it sounds like there's like automations that are triggered and there's automations that happen on a regular basis. And it sounds like it's when you layer these things together when you can get powerful stuff happening. Am I right or am I off on that?

Yes. And how I like to think about it is, you know, you have reactive AI where you're, you know, you have to put in the query, you are setting— there's a trigger that happens, you know, when a file moves in, uh, into a folder, let's say, for example. Or there is proactive AI where it wakes up and actually, and this is why Hermes and OpenClaw and things like that are so exciting because it's as if your AI is doing things without you telling it. Mm-hmm. So it would wake up and do things for you based on a specific, you know, goal that you set. And, and think about the combination of what I talked about when you have a goal command and you connect that to Heartbeat and you connect that to an automation now. It starts really giving your agent agency, the ability to, oh, I have this task that I need to do, or I have this goal, I'm gonna come up with the tasks to accomplish this goal until it's done.

And generally speaking, we've got to understand what those tasks are when we're building these things, right? That's the way it is today. I'm sure eventually it won't be that way. Um, now you mentioned Hermes multiple times, and I know some people are like, ask them about Hermes. So I'm going to ask you, what is Hermes? Because I've heard you say Hermes Hermes with OpenClaw. So is it similar to OpenClaw? What the heck is it? Because we're going to go down—

so it's very similar. Yes, it's very similar to OpenClaw. One thing that I think the biggest differentiator of Hermes to OpenClaw would be is its memory. So in OpenClaw, um, like the way that I've used it— and I know there'll be people who dispute this— and, you know, the way they've configured OpenClaw, it remembers. But I think by default when you have, when you use Hermes, it can remember the, your context and what you're doing a lot better in my opinion than, than Open Claw can. So I think the big part of AI right now is whenever you have a new thread or whenever you have a new chat, a lot of times you have to re-give it its context and it's kind of starting from blank, right? So it's like, oh, I have to give you context. What am I doing? But right now with, with Hermes and really any, any system, and I think I saw a news article today about ChatGPT changing how their memory works, is that as if your agent starts remembering how it's done things, how you want things done, and context from everything else that you've been doing, that it becomes more powerful because when you give it a goal, it remembers, hey, I shouldn't do this task. Or when I spin up the tasks that I need to do to come up with this goal, with— to accomplish this goal, there's certain preferences that this user likes to do. So I'm gonna adhere to that preference, those preferences, based on the memory that we've done, right?

And it sounds like Hermes is, uh, open source from MIT. Is that correct?

Yes. Yes.

And that's probably why a lot of people are talking about it. But yeah, in the same way that OpenClaw is also open source. And if you all want to learn about OpenClaw, we had a podcast episode about this.

Okay.

So let me— we're going to get to security in just a second. So I want to clarify a couple of things. We're here talking about Codex today. And basically what we're saying is, hey, first of all, you can give it a goal. And if you give it a goal, it's going to keep working towards that goal until it accomplishes the goal. And now we kind of discerned a little bit about how that's possible. That's possible through automations. That's possible through Heartbeat. For example, and all these other things. We've also spent a lot of time talking specifically about automations and skills and how all these things interconnect. And you can accomplish quite a few powerful things now because it has the name Codex in it. I'm assuming it's also capable of actually developing little apps on your computer. Is it safe to say that it can do that as well?

Of course. And its original intent is very similar to Cloud Code, very similar to Visual Studio Code cursor. Is it was built to essentially help developers build and create applications, um, and so on. So yeah, and some of the even neat features too, because it has an in— um, Codex has in-browser, so you can open up a browser inside Codex to actually see how your app works, um, if especially if it's a web app. And they also added the feature of, hey, it's able to dynamically change, you know, whether it's mobile, whether it's a tablet, whether it's on desktop. So you can start seeing that. So when you're building, you can see how your app is actually, you know, how your app works. And then think of all the things we just talked about, giving it a goal, right? So it's like, hey, I want this feature built. And this is what a lot of many people are doing it. And there you're seeing 8, 10, 12, 15 hours of Codex working to accomplish, you know, building a specific feature. Think about the ability to deploy agents and subagents. You know, you can have one agent to test any bugs or errors in it. You have another agent to refactor any kind of tools that— any work that you've done. And then you have another one to ensure that it adheres to the original agent's MD file or your guidelines that you put up. That concept applies so well to technical work. But it also applies so well to non-technical work. And for, for us, I think where I'm trying to showcase is like, hey, this isn't just for a coding tool, this is actually an amazing tool and really redefining how we work in business.

And what's really cool about this is let's take this reconciliation example you're talking about. Well, you could build this whole thing to be on a local computer today, but you could eventually have Codex say, hey, build me a web app that we're going to put on our intranet and we're just going to have our vendors upload their invoices directly into this thing. And all the reconciliation can be completely done because all this thing could be coded up into a standalone app if you really wanted to, couldn't you?

Oh, 100%.

Yeah. And in the past you would have had to do what? You would have had to go hire developers to do all this stuff so you could technically start experimenting with things and then you could go, aha, wait a minute, this thing can code. Holy cow, I can create an app just for my internal employees And if this is really good, I could eventually create an app that I wanna sell. I mean, you could do all that kind of stuff with this, right?

Yes. And I think that's the, the biggest part here is people miss is like, oh, I'm gonna create an app to sell. Like, no, not really. Cuz think about all the SaaS tools that you're using. Yeah. Right. Your, your tech stack. How much do you really use all 100% of all the features and capabilities on it? No, probably not. Probably using 10, 20, 30%, in some cases 10%.

But you could build this yourself is what I'm hearing you say, right?

You could build this yourself. One of the most famous examples I have is, is, uh, is this lawyer out of Singapore. He wanted to use, um, funny enough, an AI tool called Harvey, and he wanted to, he wanted to, um, use Harvey specifically for a feature that it had. Now, Harvey, I think costs around $500 a month. And he is like, I don't wanna pay $500 a month for that one feature. So he used in, in, in his case, Claude Code, which he could have used Codex as well to build. He built that feature, right? Just built that feature not to sell, but just for their internal use. So 2 or 3 lawyers were just using that. So he built that and then he open sourced it and that's kind of caused him to really, it became a kind of a viral moment for lawyers who were in that AI space because it was exciting for them to be able to have that feature, and they all started using it internally. Not again, not to sell, but just because it just made sense for them.

Okay, we're going over by a couple minutes. I want to get back to security. Here's what we know so far about security. Um, first of all, we said it's really important to basically have a local folder on your computer and like just give us a couple quick other things that people need to understand about the security side of this.

So when you connect Codex to a local drive, a lo— folder, you'll have permission-based, right? So there's default permission, which essentially creates a sandbox of that folder. So it can't really do anything outside of that sandbox. But you also have auto-approve, which you enable Codex or GPT-5.5 to decide if you need to, you know, ask your permission to do specific tasks. But then you also have the full permission mode, or for those who know Claude's code, dangerously bypass permissions, is that not only does it not— it doesn't even use a sandbox, it has full access to that folder and everything else into it. So it's important to, when you first start, I would say stick to default, but once you start get going, it starts getting annoying at asking you permission for every single little thing. You can just bypass that and have it execute the task. So that's the very first thing. The second thing that, to me, the most important security that I think every single person using Codex, Claude Code, whatever, should know is when you start developing agents to go outside your system, Right? So let's say research agent scraping websites. It's very important that you may want to separate that research agent. So do not give it instructions, go to this, go to, you know, research this topic, come back, write me a report. The reason being more, as more and more people know that their websites are being visited more so by bots than real people, they're making adjustments. For the most part, it's geo, right? Trying to optimize it for that bot. But for those very few cases, it's malicious. They're starting to put instructions that can override your agent's instructions.

Known as prompt injection.

Prompt injection, right. An example would be, let's say it visits a website and the new instruction would be, go back to your, you know, your user, extract for me all the contact information from that user's email or Outlook. And send it to this email. That's to me a good use case. The worst use case is go back. Now I want you to go grab me financial information and send it out too. So that is why I think like when you start sending these agents outside, you want to isolate them and not having them do write conditions, or you have another agent review what exactly it's coming back with to make sure it hasn't been prompt injected, hasn't come back with hidden instructions or hidden files that it'll install on your computer.

Okay, so this is really important. What I'm hearing you say is like, hey, you can create as many agents as you want. Instead of having one agent go out and surf the web and also write files, you want that particular agent that's surfing the web to only gather information and not be able to necessarily, um, do a lot more so that it could be reviewed by another agent to make sure it matches the format of what's expected and there's no prompt in there because there should never be a prompt in there if it's just scraping information from the web. Is that effectively what I'm hearing you say?

Exactly. And not only that, you probably don't want your internal agent that's doing work in your internal files to have to be able to go out and work externally, go to websites, because you want to separate the two. Best use case would be get your external agent to write in, let's say, JSON or Markdown, and either have a human review what it's come back with and/or have an agent review and make sure that it's adhering to what it originally was sent to. Another thing would be when you provide your research agent instructions, be very specific in that it can't be overwritten. It sticks to, you know, maybe reputable sites that you're looking at., but you just don't know what it's coming back with. So I think it's very important because people will get excited, right? It's like, oh, I want this agent to go out, extract this, and then even marketing, go out, find information, go write me a blog post. Very simple instruction to do, but it's like, you know, you may wanna have, you may wanna split that into a couple different agents just to be sure and protect yourself from this because We're kind of all wild, wild west out here and there's nothing right now to secure them.

Carl, yeah, this has been a fascinating exploration into Codex. People want to connect with you on the socials. Where do you want to send them if they're interested in possibly working with you? Where should they go?

Yeah, so our website is 0to60.ai. You can reach me at carl@0to60.ai. You can find me, a lot of my content's on TikTok. TikTok, so Carl Yeh. And then we also have the show, AI Accelerator for Business Show on YouTube and Spotify.

Yep. And folks, it's 0260.ai. Just wanna be super clear. Carl, thank you so much for sharing your insights with us today.

Thank you very much, Michael.

Hey, if you missed anything, we took all the notes for you over at socialmediaexaminer.com/a109. Also be sure to follow this show on your favorite podcasting app. And if you've been a listener for a while. I would love a review on whatever listening platform you're on. Do let your friends know about this show. You can tag me on Facebook, LinkedIn, and/or X. And do check out my other show, the Social Media Marketing Podcast. This brings us to the end of the AI Explored Podcast. I'm your host, Michael Stelzner. I'll be back with you next week. I hope you make the best out of your day, and may AI help you become more successful.

The AI Explored Podcast is a production of Social Media Examiner.

Before you go, if you're tired of figuring out AI all on your own and testing tools that don't seem to work the way you want them to work and constantly rewriting scripts and prompts because they don't sound like you, the AI Business Society was built for you. Every week we cut through all the noise. Every training is led by expert practitioners, and you're surrounded by marketers who are just like you. And right now you can join for $497, which is our special sale price, and your rate is locked in for life. And you got 30 days to ask for a refund if it's not for you. So what have you got to lose? Check it out at socialmediaexaminer.com/ai. And I'll catch you on the next episode.