The argument in one line.
OpenAI is collapsing the distinction between chatbot and coding agent into one platform, and when Codex’s full power reaches ChatGPT’s billion users, every role-specific plugin becomes a potential startup killer overnight.
Read if. Skip if.
READ IF YOU ARE…
- You build products that could become OpenAI plugins or are evaluating whether to develop on top of the OpenAI ecosystem.
- You work in enterprise software and want to understand how domain-specific agent plugins are displacing specialized SaaS tools.
- You use vibe coding platforms like Lovable or Replit and want to understand the competitive threat from OpenAI Sites.
- You follow the AI agent landscape and want a practitioner take on what the Codex-ChatGPT merger means for day-to-day workflows.
SKIP IF…
- You want deep technical implementation details — this is strategic commentary, not a coding tutorial.
- You already follow OpenAI announcements closely; most of the primary source footage will be familiar.
The full version, fast.
OpenAI confirmed that Codex is merging into ChatGPT to form one intelligent platform. Three new pillars: role-specific agent plugins (data analytics, creative production, sales, finance, product design, investment banking) that bundle company context, domain knowledge, and collaborative output; Annotations that let users give inline direction on any AI-produced artifact; and Sites, which lets you deploy any Codex output as a live hosted web experience from a single prompt. The strategic implication: once Codex capabilities reach ChatGPT’s roughly one billion users, any official plugin immediately has the largest distribution in software.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:24
01 · Introduction
Split-screen of Codex and ChatGPT; teases the merger announcement.
00:24 – 02:29
02 · Codex and ChatGPT are merging
OpenAI confirms unified platform; agents on desktop, mobile, browser, and inside tools like Excel and Slack.
02:29 – 08:50
03 · OpenAI plan for agents
Evolution from conversations to goals; /Goal command; proactive agent behavior; chorus.com sponsor; OpenClaw heartbeat analogy.
08:50 – 11:26
04 · Role-specific Codex plugins
Six enterprise plugins: data analytics, creative production, product design, sales, public equity investing, investment banking. Each has context, domain knowledge, and collaborative output.
11:26 – 15:19
05 · Why these are startup killers
Harvey (AI for law) as canonical example. Plugins do what entire startups were built to do. ChatGPT distribution makes the threat existential.
15:19 – 17:15
06 · AI is barrier destroying
Commentary on why splitting Cowork and Claude Code is the wrong direction; AI should erase the technical-vs-nontechnical barrier.
17:15 – 20:00
07 · Annotations inside Codex
New feature: select any part of a Codex artifact and annotate inline. Demonstrated on a personal site and a spreadsheet.
20:00 – 23:06
08 · Codex Vibe Coding — Sites
Sites lets you publish any Codex output as a live shareable web experience. Riley demos a personal landing page auto-updated weekly by an agent. Direct threat to Replit and Lovable.
23:06 – 26:18
09 · Final thoughts
Recap of all three pillars. Prediction that Sites grows into a full vibe coding platform with auth, DB, and AI features. CTA to follow.
Lines worth screenshotting.
- OpenAI merging Codex into ChatGPT means roughly one billion ChatGPT users will eventually access every Codex plugin without switching apps.
- Role-specific agent plugins are startup killers because they package what entire companies like Harvey were built to deliver into a single toggle.
- The /Goal command shifts Codex from reactive to persistent — the agent works until the outcome is achieved, not just until the prompt is answered.
- AI deepest value is erasing the technical-versus-nontechnical divide; platforms that reinforce that divide will lose to ones that collapse it.
- Sites turns Codex into a direct competitor for Lovable and Replit; the missing pieces are authentication and a managed database, which are likely coming.
- Being an official OpenAI plugin becomes an enormous distribution advantage the moment ChatGPT full user base gets plugin access.
- Agents running persistently in the cloud are a different category from tools you visit — they are always-on teammates that act without prompting.
- Context over email, calendar, and messages is what enables AI to move from reactive to proactive; granting that access is granting the AI your goals.
- Annotations close the last-mile gap between near-final AI output and polished deliverable without forcing the user out of their working context.
- The three things every enterprise plugin shares — company context, domain knowledge, collaborative output — are the exact three things specialized SaaS startups are built to provide.
What the Codex-ChatGPT merger means for builders.
WHAT TO LEARN
The gap between a chatbot and a coding agent is closing, and the moment it closes for a billion users, the distribution math changes for every product built on top.
- Codex merging into ChatGPT is a distribution event, not just a product update — every capability Codex has will eventually reach all ChatGPT users without any change on the user part.
- Role-specific agent plugins combine company context, domain knowledge, and collaborative output — the same three things specialized SaaS startups are built to provide, which is why they displace them.
- The /Goal command represents a shift in how you use AI: instead of prompting for steps, you describe an outcome and the agent runs until it gets there, requiring much less hand-holding.
- Proactive agents require deep access to your calendar, email, and messages; that access is now the product differentiator, not just the model.
- Annotations solve the last-mile problem: once an agent produces a near-final artifact, you can give it inline direction on any part of it without starting over, keeping the human in the loop.
- Sites compresses the distance between having an idea and having it live on the internet to a single prompt; once auth and a database are included, standalone vibe coding platforms lose their core value proposition.
- Platforms that separate agentic coding from general AI tasks create friction that users avoid — the winning approach is one unified surface where any task is available without the user choosing a mode.
- Being an official plugin on OpenAI platform may become a distribution channel as significant as being in an app store, especially as agents begin selecting plugins autonomously without user direction.
Terms worth knowing.
- Codex
- OpenAI agentic coding and task platform, separate from ChatGPT, that can run long-horizon tasks, execute code, and connect to enterprise systems. Being merged into ChatGPT.
- /Goal command
- A Codex feature where you state a final outcome and the agent works autonomously until it is achieved, rather than responding to a single prompt.
- Agent plugins
- Role-specific extensions for Codex that bundle company context, domain expertise, and tools for a particular job function such as data analyst or creative director.
- Annotations
- An OpenAI Codex feature allowing users to select any part of an AI-generated artifact and give the agent inline instructions to revise it.
- Sites
- A new Codex feature in early preview that lets users publish any Codex output as a live, shareable, interactive web experience with a single command.
- Harvey
- A legal-AI startup that compiled deep domain knowledge about law into a product. Used as a benchmark for single-domain SaaS that enterprise plugins can replace.
- Chorus
- Riley Brown own product, which lets users add Claude Code or Codex-style AI agents to iMessage, WhatsApp, and Slack group chats.
- OpenClaw
- A third-party open-source agent framework mentioned as an early example of proactive agent behavior, including a heartbeat that checks context every 30 minutes.
- Vibe coding
- Informal term for building software by describing what you want in natural language and letting an AI agent generate, run, and deploy the code.
Things they pointed at.
Lines you could clip.
00:06
“Their giant plan is to combine both of these applications. This is not a conspiracy theory.”
11:11
“These plug ins do what an entire startup does.”
15:08
“The coolest part about AI is its barrier destroyed.”
24:42
“When you can make sharing a site as easy as sharing a doc, it transforms how expressive we and our agents can be at work.”
24:22
“Anyone on your team can create software through just a single prompt, deploy it across the company, and then ask agents to autonomously maintain that output.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogystory
00:00So this here on the left is OpenAI's Codex, and this here on the right is OpenAI's ChatGPT.
00:06And what if I told you that their giant plan is to combine both of these application? This is not a conspiracy theory. I announced it today in this presentation that outlines the entire future of OpenAI, and it outlines exactly where AI agents are going.
00:23I'm very excited to share. In the next few weeks, we're going to be putting Codex into ChatBet. It's that simple.
00:30With Codex and ChatBet, everyone will have access to the most powerful agents at work and in daily life. Your agents will be available in one unified experience wherever you're working, whether that's on your desktop, on your phone, or in your browser, and in the tools you know and love, like Excel or Slack or even our top enterprise feature requests, even in PowerPoint.
00:51Okay. Let's break this down real quick. Quick.
00:52So as of today, codex and ChatGPT
00:55are different products. However, in the very near future in the very near future, these two products are coming together. And, yes, if you've been following my videos, you do know that there's actually a codex feature inside the ChatGPT app.
01:10This codex feature allows you to connect to codex as kind of this remote.
01:15I can go for a walk, go on with my day. I mean, I do this often. I do this every day where I just open up Chat GPT and I'm using codecs through ChatGPT, but they aren't fully integrated.
01:26It's not like a seamless product experience yet. And so over the coming months, we're gonna see these tools really come together, whether it's on a desktop app, your mobile app, or your browser. And speaking of desktop apps, like, I actually have a ChatGPT desktop app on my computer.
01:44I don't use it. I don't even know if OpenAI is using it. I use the Codex app, and the fact that they're separate apps is still a little bit confusing.
01:52And so one of their main beliefs is that, like, you shouldn't have to decide which power tool to use. Their whole goal is to just put intelligence in one single intelligent platform and then put that everywhere you work. To make all this possible, our agents are gonna to evolve quite significantly
02:08from primarily running locally on your computer to running persistently, working persistently in the cloud. So you might not be working twenty four seven, but your agents in the cloud will be working twenty four seven.
02:18In summary, our vision for putting coex into ChatHPT is agents wherever you need them, constantly accelerating work in the background, and helping everyone get more done. But today, I also wanna share something else. I wanna share how we're going to elevate the work that you do with agents, making all your agents smarter and more effective.
02:33So work has already moved to asking agents to accomplish tasks. This is not anything new, but we're now heading towards this next level of autonomy where you give an agent a goal, you let it figure out how to get there, and then you have it keep working until it succeeds at that goal. And if you've been following our latest releases with Codex, we showed features recently like the goal command, which really bring this to life.
02:52Today, we're taking another step forward. So we're gonna show you a few things. First, we're going to show you role specific agent plugins that give agents the domain knowledge and tools to take on meaningful work for you and your team.
02:59Second, annotations, a new way to collaborate with the model directly in the tools you use every day from spreadsheets to presentations. And third, something new but familiar, a way to turn work into something interactive, shareable, and much closer to software. So he mentioned conversations
03:12to goals. And in the Codex app, if you, uh, haven't been watching super closely, they released a feature, uh, called slash goal.
03:19So if you're using Codex and you type slash goal, this pops up right here.
03:25And basically, it allows you to just type in any sort of final output or final output goal that you want and codex will just work until that's done. But beyond goal, which is just like super long lasting tasks, there's also something else that OpenAI has been talking about a lot, specifically Sam Altman.
03:45In one of his recent interviews, I actually couldn't find it. I tried to find it, but he starts talking about something that's pretty interesting that just things should start to happen around you. And in order for AI to do things for you in the background, it actually has to know your goals.
04:02Right? And that's why context is important. So AI is starting to get access to your email.
04:08Right? You're giving it your email. You're giving it your calendar.
04:12And in some cases, you're giving it your iMessages and text. On Codex, it can actually fully control your iMessages and it can text people directly through Codex, which I find to be incredibly interesting and fun.
04:26But through this, they can actually understand your goals. And so when AI understands your goals and the things that you're trying to do and it's very well connected and can make high conviction guesses as to what your goals are, it's able to spin up agents in the background that might help you accomplish those goals.
04:46And that's one thing that OpenClaw did really well is your agent started doing things that surprised you based on your goals. Right?
04:56This is the feature called heartbeat through OpenClaw was basically like every thirty minutes it's gonna wake up, it's gonna check your stuff. Hey guys, quick break to talk about Billy, an AI agent that I created on chorus.com.
05:10We started building this because we wanted an agent with the reliability of Codex and Clawd Code, the customizability of OpenClaw and Hermes, and we wanted people to be able to set it up in three minutes on iMessage. So Billy, my Chorus agent sends me a morning update along with an updated mini app which shows how many partners at pitches, how many things are waiting on me, and how many meetings I have in that day.
05:32My favorite skill is it comes pre baked with the ability to search YouTube. So I could say something like, hey, buddy.
05:39I need you to research my last three videos and tell me how I can do better. Look at my competitors and tell me what YouTube videos that I should make that would be really popular that my audience would really like. Think for a while, ten minutes, and do in-depth research and prepare a PDF.
05:55And so because this is an AI agent just like OpenClaw or ClaudeCode that runs persistently on a computer, the agent can do anything that ClaudeCode can do.
06:05On chorus.com, you can see all of its files including the agents dot m d identity and soul if you're used to OpenClaw. You can configure skills.
06:14There's a marketplace of hundreds of skills that you can get. There's many different platforms. You can connect it to iMessage, WhatsApp, and Slack, and you can set up any of these connections.
06:24You can also set up automations incredibly easily just by typing to your agent. And my favorite feature is you can add your agent to a group chat. So say, hey, Billy.
06:37This is Anj. I'm filming a video.
06:43Give me some ideas for use cases for the agent.
06:48And so this will create a new group chat and my agent, Billy, will immediately start responding. And there you go. Here's Billy's response.
06:58Yo, Ange. Billy here. You are now to act as Ange's and I's executive assistant and help us with everything we need help with.
07:17Yep. So this is basically clog code running in your group message directly from iMessage.
07:22This is on chorus.com. We've been working on this for about two months and it's really starting to come together. I would love your feedback.
07:29Let me know how it goes. The link is in the description. Your email calendar, iMessages, whatever it is that you give it, maybe you're linear or Slack.
07:36And it's just gonna try and do something useful. And I remember when I first used OpenClaw, it actually generated a presentation about one of the candidates that I was going to interview for a really important role, and it was actually really useful.
07:48And it just texted it to me. And I was like, oh, that's where AI agents are going.
07:53And so this is just kind of this proactive future, proactivity rather than reactivity.
08:00Codex is an incredibly powerful tool as it exists right now. Codecs is incredibly powerful, but you still have to go to the platform.
08:08And even with automations, you have to set up the automations. Soon, in this one intelligent platform future, the AI might just start doing things for you.
08:19It'll have enough context with all of your tools that it'll just start doing things for you. It'll just be like, yep.
08:25Hey. I made this for you, by the way. Like, I hope this is useful.
08:27Let me know if I should keep doing this.
08:29And that's kind of the future we're going to. Anyway, let's get back to the video. Imagine having agents that empower your best marketer.
08:36Agents that can multiply the work of a sales rep or that can build dashboards for a seasoned data analyst. With every specific business role, there are different tools, different context, different domain knowledge, and skills that you need to know. So how do we make this effortless for you, and how do we make this effortless for everyone at your company?
08:50Today, we're introducing six role specific plug ins. They're rolling out in codex, and they're gonna be coming into ChatBit soon. These enable role specific work, packaging skills, workflows, and connections to your business systems like Salesforce, Databricks, Snowflake, and more.
09:03Okay. So you notice here that they have role specific plugins. So we have sales data analytics, creative production, product design, public equity investing, and investment banking.
09:14Let's head over to Codex really quickly. I did notice that there are some new plugins in here such as product design.
09:23I haven't seen this before and I'm pretty sure that this will allow you to just create designs
09:28and they'll render right here on the right side of the screen. Skills, workflows, and connections to your business systems like Salesforce, Databricks, Snowflake, and more. I like to think of them like a teammate that's already gone through company onboarding and is just waiting to get to work.
09:41It's it's a little bored, and it's eager to get started to do the work. So let me share a few examples.
09:46We'll start with data analytics. I'll play the video in a second, but just to call out what we're gonna see here. When you give an agent a task, it's going to bring three things together in Codex.
09:54First, the agent's going to have the context it needs. It's going to pull from our data, from everything from across organization, the apps, and the enterprise software we use. Second, it has domain knowledge.
10:04It knows how we like to approach the problem. It knows how we like to write our summaries. It knows how we like to propose next steps in action.
10:10Third, the work is collaborative. Everything it produces will be editable, so I can go in, refine it, and keep working with the agent before we hand it off to the rest of the team. Let's say I have a problem to solve, which is I've noticed a spike in cancellations.
10:20So I'm gonna ask Codex to take a look. And in just a few minutes, this is Fedel. It's going to understand our semantic data layer, write a query, and produce this report.
10:28I can go in. Maybe I prefer this as a bar chart. Awesome.
10:30That looks better. Perhaps I want to verify the query they wrote under the hood, so I can go in, verify the query. Great.
10:36And maybe that gives me an idea for a different title for this chart. So, great, I can customize the title, and maybe now I'm ready to share. Now if we stopped here, I still have a lot of work to do.
10:43I have to massage this output into some other output that I'm gonna give to the rest of the team. But here, have a few options, and let's ask the agent to produce a deck. So the agent will get to work.
10:50And in just a few minutes, I'm going to have a deck that I can share with the rest of my team in the format that we know, understanding why is there, uh, spiking cancellations and what are recommended next steps. Okay. So they're announcing these enterprise
11:00plug ins, and all of these plug ins have three things. Right? It has context over everything that's going on at your company, what we just talked about.
11:08It also has deep domain knowledge. And so, uh, there's a startup that many people talk about called Harvey. And so Harvey is like chat GBT for law.
11:19And they basically compiled all of this domain knowledge about law and they created this product called Harvey. And that's why everyone calls these types of products that come out startup killers because these plug ins do what an entire startup does.
11:35Right? In this case, Harvey. And then the third thing is that it's collaborative.
11:39And we'll get to this more later because OpenAI also announced their new vibe coding tool which allows you to create any sort of internal tool and immediately share it with your team. And so these are like little internal apps that you can immediately share with your team and all of these are collaborative.
11:54And so that's what all of these plug ins have, context domain knowledge, and the ability to collaborate.
12:03Business problem with a lot of ambiguity to a query that I can just throw off while I'm thinking about something else and ultimately get to a polished deck with analysis and fixes. So now I'm gonna show you another example. Let's look at creative production.
12:13Every company has people who are more science focused and people who are more art focused. And these people aren't always the same. And trust me on this, was banned from art class in primary school.
12:22But with Codex, I can work with a creative production plugin to start creating on brand assets. So here, I'm going to drop in an image of the products or inspiration, and we're gonna get started. Now the way that our team likes to work is we like to start with the moon board.
12:34So Codex is asking you a few questions for what to brainstorm, and now we have a bunch of options. And this step is really important because it allows our team to really exercise our human taste and judgment. Now perhaps I like this image.
12:43I want to customize it a little bit. Let's tweak the image so that it's, like, one second earlier where the goal is still being placed. Awesome.
12:49Okay. Much better. With this plugin, we spent a lot of time figuring out how to really pull the taste and intuition out of everyone, out of the people using the product.
12:55So here's another feature, which is remix. I can change this to make a more dramatic look with the counter. Awesome.
13:00So now we have a couple images. The next step we take here is we want to turn this into an ad. So what I can do is I can go to Codex and say, hey.
13:08Here are a few images I like. Please create an ad for me. And using ImageGen, in just a few minutes, I have ads in a variety of formats that we can use to produce an ad.
13:16Now, again, this is not the final form factor. What I really need to hand off is I need to put this in either Figma or Canva, and I need to create a reusable asset for my marketing team so that they can wire this up. So, again, with just a few clicks, we produce these assets, put it into Canva, and now now we have a reusable editable artifact.
13:29OpenAI is taking all of their knowledge that they've gained over the last two years, and they've probably watched these companies closely.
13:35And they're like, yep. This could be easily be a plug in or a feature inside Codex, which is their new super app.
13:43But, like, now we know that OpenAI ChatGPT and OpenAI Codex are gonna become one thing.
13:49So that means every single ChatGPT user will be able to use their AI ad maker, which is a plugin. And so these plugins are gonna start out in the enterprise, but over time they're gonna work their way down to every user.
14:03If someone asks for an ad, it'll just generate ads using the right plugin. You won't even have to ask for it. Right?
14:10That's the direction that they're trying to go to. If they truly wanna create this one intelligent platform, it should just be able to use plugins. Right?
14:17The agent's just gonna be like, yep. I can use this these plugins. He doesn't even have to ask me to do so.
14:22I'm just gonna do it. And so all ChatGPT users are gonna eventually get access
14:27to these new plugins. Again, in just a few minutes, I created a whole new asset library going through the flow that our team likes, which starts brainstorming and then refinement and then reusable asset. Previously, this would have taken a creative team days or weeks, but now anyone, even me, can take a stab at this if we're interested in creating ideas.
14:43So Codex is a tool, a single tool, that your whole company can use to do any work you can do on a computer. And, critically, it's not just that the data analyst can use the data analytics plugin or a marketer can use, uh, the creative production plugin. Actually, anyone at the company can try on any of this work.
14:59Did you hear that? He said anyone at the company can try any of this work and this is the coolest part about AI to me that a lot of people
15:08are actually concerned about for good reason. And this is the main reason why I hate that Anthropic made a distinction between co work and Claude Code.
15:18The coolest part about AI is its . Right? There's now less of a bit of barrier between someone who is like technical and someone who is nontechnical.
15:29Right? Someone who is technical may want to be able to do some marketing work. And because AI has gotten so good, it creates like a baseline of of skill that you just have access to.
15:39Right? People just have access to intelligence. And so you should be able to just go to your AI and fire off a prompt.
15:45You shouldn't have to learn how to use co work that's separate from Claude code. And, uh, on the flip side, if you're completely nontechnical, you should be able to talk to your AI agent and say, hey, I wanna build a production ready application and I wanna run it locally and I wanna be able to put it on the Internet. And the AI agent should just be able to do that.
16:03Creating these barriers between coworking tasks or general agent tasks and coding tasks to me is a little bit annoying, and I don't think this is where it's all going. In fact, I made the prediction publicly that Claude or Anthropic is going to remove coworking because having the distinction between the two is unnecessary and honestly annoying.
16:25I've talked to many people and no one likes that there's co work unless it's the only thing that you've learned. But as soon as you wanna upskill and learn Claude code, then you're gonna run into some, like, issues.
16:36And so that I think is something that OpenAI realized pretty early on that anyone should be able to just do anything. So if you have an idea, you can just do it and that's kind of what AI allows you to do. And that's what makes it so damn cool.
16:50Let's keep going. So those are just two plugins. We believe in an open ecosystem.
16:55So in addition to the six role specific plugins rolling out today, we have more than a 100 plugins available for Codex. Our vision is that AI works with the software your teams already use, bringing in more context and intelligence into every workflow. So now let's talk about another feature another feature shipping today.
17:08As we saw, agents are producing more and more near final work. But there is this last mile problem. How do we collaborate with the model in getting to that final work?
17:15Today, rolling out annotations. Now when Codex produces a deliverable, you can select anything on the deliverable and ask Codex to explain or edit. You could already do this with apps or sites that Codex created.
17:24Now you can do it with any asset. So here I have a spreadsheet. I've asked Codex to do quite a robust analysis with several worksheets.
17:29But if I want to look at this data specifically in a chart format, I can select it and just ask Codex to explain or ask Codex, in this case, to produce a chart. So let me explain what he means by this. So whenever you're using Codex,
17:39they're super app to do any vibe coding where you ask it to create an app and it creates an app and it shows you this browser, you've been able to annotate. And so I could come here and I could say change this, make it shorter.
17:55And I could enter this as an annotation right here. I could add multiple annotations.
18:01I could say these icons should be bigger, and I could enter this in. And now it sees two annotations.
18:09Now I can say, please make these changes. And this will tell the AI exactly what to do. And these annotations gives the agent context.
18:18I actually don't wanna make those changes because this is my personal site, so I'm not gonna make them. But the what he just said in the video is something that's really interesting.
18:27So here I had AI create a spreadsheet. So here's the spreadsheet of Microsoft Alphabet and Salesforce financial report. Apparently, I've actually never tested this.
18:37I can actually come in here and highlight a certain part of this, and I believe I can annotate. Oh, I can annotate directly on the spreadsheet.
18:47And I could say create a chart for this.
18:52And I could enter this in and I can annotate directly on any of these artifacts, which is really cool.
19:01You're you can now annotate on any of these artifacts, whether they're spreadsheets, documents, or presentations.
19:07This is incredibly powerful because it gets us from this from Codex producing a near final artifact to keeping you in flow, collaborating with the agents so you can produce higher quality work faster. Now you can produce this higher quality work faster in the common formats that we all use at work today. This is obviously super important.
19:22But we have one more thing to share, which is a new way to share intelligence. And this is a powerful but familiar format. So when today, when AI helps you create something valuable, you usually have to squeeze that into a document or a spreadsheet or a slide deck.
19:35And these formats are useful, but they're optimized for authoring and presentation, not for understanding. So what if we didn't have to?
19:41What if we could shape information around the best format for consumption that is as flexible and interactive as we want? That is why we are introducing Sites in early preview. With Sites, you can take anything you create in Codex and turn it into a secure, shareable experience, whether that's a dashboard, a prototype, a tool, or even a full app.
20:00Yep. You heard that right. They are coming after Lovable and Replit.
20:05And, again, I know he's it's it's an early preview,
20:08and these are just small shareable sites that you can share with your team. But over time, I guarantee you this is gonna grow into basically fully hosted production ready applications over time.
20:19And you can now vibe code directly in Codex and and it will host them for you. This was probably the biggest announcement today. So instead of sending a doc or a spreadsheet, you can just publish the work securely to your team and let people in your workspace interact with it directly.
20:33Now I said familiar because in a way, is an obvious solution. The software engineers on our team have been building microsites for years, but it wasn't easy for everyone. It involved cumbersome setup, and so most people weren't doing this most of the time.
20:43But when you can make sharing a site as easy as sharing a doc, it transforms how expressive we and our agents can be at work. We've been using sites for a while now inside OpenAI, and they are just taking over as a format that we share. So I'll give a few examples.
20:55One example is you can take a financial forecast and turn it into a real time business dashboard for the whole org. Or another, you can take a product plan and turn it into a working prototype that your team can review together. This is something we actually do all the time.
21:05Or for this event, in fact, our marketing team created a site for the event to manage the stream audience and to track follow-up. So now anyone on your team can create software through just a single prompt, deploy it across the company, and then ask agents to autonomously maintain that output with the full business context, domain knowledge, and all the plugins that you you're using at your company.
21:23That is really, really important. And so one thing that I wanna show you, uh, kind of how I've been using Codex
21:28is I've been using Codex, uh, to create this personal landing page. And you'll see here this personal landing page has real time followers of these, uh, channels.
21:39It also has weekly agent updates. So here you can see weekly agent updates. Every single Friday, there's an automation that will automatically trigger this agent to update this and then redeploy it to rileybrown.xyz.
21:54And so for this, I'm using Vercel. So right now I have to use at Vercel. And Vercel is the hosting service that I use.
22:02That's where it's hosted so that it can be on the Internet. OpenAI is releasing one when all you have to do is press at sites. I do see that I no longer have access.
22:13I had access earlier. Now I don't have access. I think they're still trying to figure out how they're gonna roll it out.
22:19But all you have to do is type at sites, create an an a little app, or you can create a real time dashboard with your company's data, and it will stay updated.
22:29And so this right here is the future of AI agents. Right? You can create any interface you want right here on the right side.
22:36Right? Whether it is an HTML file, whether it is a fully interactive site, whether it's an immersive game that allows you to, like, play in a three d world in order to learn a certain subject, you can or you will be able to very soon host these and share them with your team. And your team may even be able to edit them.
22:55And so this is really, really cool. You can fully vibe code apps directly inside Codecs after their latest release. Sites are rolling out in preview to business and enterprise teams today, and they'll be coming to other plans soon.
23:06So just to recap, agent plugins built for specific roles that can give you a goal that you can give a goal and they that where the agent can carry out complex work on your behalf. Second, annotations that change how you collaborate with models to produce stronger content and insights. So here's a good summary of what he talked about and where this is all going.
23:21They're gonna have agent plugins, and all of these agent plugins are going to have expert domain knowledge. They're gonna be connected to all of your different tools. They're gonna have context, and they're gonna be collaborative.
23:32So you can create some sort of deliverable at the end, share them with your team, and be able to edit them together and work together. And then with all of these artifacts that get created for knowledge work, whether it's a sheet, a deck, or some of one of these sites, right, which is basically any sort of app.
23:49Right? So in Codex, you can create these little mini sites that can be any format. You can annotate directly on them.
23:56And so they're they're really trying to push teams and being able to work on these deliverables as a team. And these annotations allow you to tell the agent exactly what you want it to do. And it's almost like commenting on a document if you've ever used, you know, Canva or Figma or even Google Docs.
24:12You can comment and this will tell the author what they need to, uh, what they need to fix. It's like that but for AI agents. And so annotations are really cool.
24:20And then, of course, sites, I believe, will evolve into a full vibe coding platform because the value prop of Replit and Lovable are that they host your site, they give you, um, authentication, they give you database, and they give you access to AI features.
24:38All of those things might be added into this site's feature over time. I highly anticipate it will because I know that, uh, Lovable and Replit are their biggest clients.
24:48And they see this as a massive opportunity to just insert it into the platform. And this is why I believe that vibe coding platforms, it'll be really hard for them to exist in the future when it's just a tool that codecs can use. And so all of these new updates to Codecs are coming very soon.
25:08And then the biggest update is that Codecs is basically going to be inserted into ChatGPT. Now how are they going to insert it into ChatGPT?
25:17That's still unknown. We don't know how that's going to work. We don't know when that's going to happen, but we do know that the power of Codecs will be in the hands of everyone who uses ChatGPT.
25:28You can get your app to show up inside Codecs. And so that means the total addressable market right now is tiny compared to what it will be in the future. Soon, I don't know what what is it?
25:41A billion users on ChatGPT are going to have access to these plugins. And that is the value of being an official plugin on OpenAI's ecosystem because that means anyone could theoretically use this Remotion plugin or your Canva plugin.
25:57OpenAI is combining Codex and ChatGPT. They're adding a bunch of new features, specialized agents, annotations, sites.
26:06These are all gonna come to fruition, and I think all of the other big super app companies are going to follow. And, of course, I'll be keeping you updated right here. So make sure to follow for that.
26:16I'll see you here in the next video.
The Hook
The bait, then the rug-pull.
OpenAI just publicly confirmed what the product roadmap had been hinting at for months: Codex and ChatGPT are merging into a single intelligent platform. Riley Brown watched the Intelligence at Work event so you do not have to, and his breakdown covers not just the what but the competitive implications — specifically why role-specific agent plugins are startup killers and why Sites is the opening shot at Replit and Lovable.
Named ideas worth stealing.
11:00list
Enterprise Plugin Formula
- Context (company data + org systems)
- Domain Knowledge (role-specific expertise)
- Collaborative output (editable, shareable with team)
Every effective enterprise agent plugin must combine these three things. Missing any one breaks the loop.
Steal forEvaluating any SaaS-to-agent transition or building a plugin on OpenAI platform
03:00model
Conversations to Goals evolution
- Conversations (one-shot prompts)
- Tasks (multi-step with human approval)
- Goals (/Goal — run until done)
- Proactive background agents (no prompt needed)
The progression from chatbot to autonomous agent in four stages. Each stage requires more context and trust.
Steal forPositioning any AI product on the autonomy spectrum
How they asked for the click.
VERBAL ASK
25:46subscribe
“I will be keeping you updated right here. So make sure to follow for that.”
Clean close tied to ongoing coverage promise; low pressure, high continuity.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
Visual structure at a glance.
