Modern Creator
Why Modern Creator?
Riley Brown · YouTube

OpenAI Merges ChatGPT and Codex

Riley Brown and Ras Mic dig into GPT-5.6, Codex's background computer-use, and why self-scoring agent loops are turning coding tools into a general operating system.

Posted
yesterday
Duration
Format
Interview
educational
Views
13.6K
603 likes
Big Idea

The argument in one line.

OpenAI merged ChatGPT and Codex into one agent-native app and shipped GPT-5.6, but the real shift isn't the model bump — it's that background computer-use, self-scoring loops, and parallel agent threads now make Codex usable as an everyday operating system for developers and marketers alike.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…
  • You already use Codex, Claude Code, or a similar coding agent and want concrete workflows (loops, multi-threading, computer-use QA) beyond just prompting.
  • You're a marketer or founder curious how agentic AI applies outside of writing code — ad scripts, thumbnails, video analysis.
  • You're deciding whether to keep paying for multiple AI subscriptions and want a read on how GPT-5.6 actually compares to the top-tier model on real tasks.
  • You want to understand OpenAI's product strategy (merging ChatGPT and Codex, sunsetting Atlas) and what it signals about where AI tools are heading.
SKIP IF…
  • You want a technical deep-dive on GPT-5.6's architecture or benchmarks — this is a hands-on workflow conversation, not a model-card breakdown.
  • You're looking for beginner AI basics — both hosts assume familiarity with coding agents, MCP, and agent terminology.
TL;DR

The full version, fast.

OpenAI merged ChatGPT and Codex into a single app and released GPT-5.6 (Sol, Terra, Luna), with Sol aimed at long-running agentic work. The hosts argue 5.6 is a strong, cheap version bump but not a true rival to the top-tier flagship model on hard tasks. The real value is four workflows: background computer-use that QA-tests apps while you keep working, self-scoring loops (define success criteria plus a feedback set, then let the agent iterate until it clears the bar), picking the right pre-built tool stack instead of building from scratch, and running multiple parallel agent threads so each task gets a fresh, uncluttered context. New Sites and an improved in-app browser turn Codex into a full vibe-coding platform, and OpenAI is folding its Atlas browser team into that effort.

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 →
Chapters

Where the time goes.

00:00 – 00:40

01 · Introduction

Riley Brown cold-opens on the ChatGPT/Codex merge and the new GPT-5.6 models, then brings on Ras Mic to break it down.

00:40 – 02:20

02 · OpenAI Merging ChatGPT and Codex

The new 'Work' tab is explained as OpenAI's play to expose ChatGPT's ~983M users to Codex-level agentic capability they don't know exists.

02:20 – 05:05

03 · Sol, Terra, Luna — the new models

Sol is for long-running agentic work; Terra and Luns will likely live in the consumer Work tab. OpenAI is praised for token efficiency and tool-calling discipline versus Gemini's 'psychosis' spiral.

05:05 – 08:20

04 · The Replit/Lovable benchmark

A one-shot 'build a Lovable clone' test: the top-tier flagship model did it in a four-sentence prompt; GPT-5.6 needed a longer prompt to succeed, reinforcing it's a strong-but-not-flagship-class model.

08:20 – 09:00

05 · Pricing of 5.6

GPT-5.6 holds the same price as 5.5 while being meaningfully more capable — framed as OpenAI building efficiency ahead of a future flagship-class GPT-6.

09:00 – 15:10

06 · Computer use

Codex's computer-use feature is demoed testing an app end-to-end in the background; Ras Mic describes running it via a mini PC and using it as an always-on QA tester.

15:10 – 16:15

07 · Record and replay for knowledge work

A new feature records up to 30 minutes of screen activity and converts it into a reusable agent skill, aimed at non-developers automating repetitive tasks.

16:15 – 20:20

08 · Loops

Ras Mic defines the agent loop — the agent reviews and scores its own code changes, then keeps fixing until it hits a self-assigned 5/5 — and shows scheduling a recurring PR-review automation.

20:20 – 26:10

09 · Loops for marketing

Riley extends the loop pattern to marketing: scraping top-performing competitor ads via Foreplay and having Codex grade its own generated ad scripts and thumbnails against them.

26:10 – 28:20

10 · Plugins and social media scraping

Scrape Creators is used to pull competitor videos from any platform; combined with a Gemini video-ingestion skill, Codex can watch and analyze video content to benchmark quality.

28:20 – 30:45

11 · Tools

Picking the right tool for the job (Blender's MCP for 3D, Gemini for video ingestion, Convex for the database layer) is framed as the difference-maker once computer use and loops are in place.

30:45 – 34:05

12 · Building block economy

Referencing Vercel's Guillermo Rauch, the hosts describe agents increasingly assembling pre-built services (Convex, Eve, Daytona) like Lego pieces instead of writing infrastructure from scratch.

34:05 – 37:20

13 · Use multiple threads

Running features in separate Codex threads gives each task a fresh, uncluttered context; ten threads were spun up from one conversation to build and test ten features unattended.

37:20 – 40:00

14 · New Sites feature

A live demo of typing '@sites' into a new Codex chat to generate and host a working website in one prompt, turning Codex into a full vibe-coding platform.

40:00 – 41:50

15 · Updated browsers

Codex's in-app browser now supports multiple tabs; OpenAI is sunsetting its standalone Atlas browser and merging that team's work into Codex.

41:50 – 47:51

16 · Agent-native operating system

The conversation closes on Codex as an early AI operating system, the coming design/writing skill gap, and the claim that the education gap on these tools is bigger than the tool gap itself.

Atomic Insights

Lines worth screenshotting.

  • OpenAI merged ChatGPT and Codex into one app and released three new models — GPT-5.6 Sol, Terra, and Luna — with Sol positioned as the flagship for long-running agentic tasks.
  • GPT-5.6 is framed as a version bump, not a generational leap; the hosts expect the real jump to arrive with GPT-6, the way Opus 4.5 was the last big point release before a bigger shift.
  • Terra and Luna are expected to quietly power the new consumer 'Work' tab, while Sol serves developers and power users running long agentic tasks.
  • On a one-shot 'build a Lovable clone' benchmark, the top-tier comparison model built the whole app from a four-sentence prompt, while GPT-5.6 needed a longer, more detailed prompt to succeed.
  • Gemini scored well on raw intelligence but performed poorly as an agent in this test, spiraling into 15-20 redundant tool calls on tasks GPT-5.6 finished efficiently.
  • Codex's computer-use runs in the background, letting a user keep working in another window while the agent drives its own browser session — unlike some coding agents that visibly take over the whole screen.
  • A basic agent loop has three parts: a well-defined prompt, a feedback mechanism (competitor ads, reference thumbnails, or a working app to test), and a stated success criterion, after which the agent repeats and self-scores until it clears the bar.
  • Marketers can run the same loop pattern developers use for code review: scrape a set of top-performing competitor ads or thumbnails, then have the agent grade its own drafts against that set until it matches quality.
  • Record-and-replay converts a screen-recorded task into a reusable agent skill, replacing the need to hand-write a precise prompt or SOP for repetitive knowledge-work tasks.
  • Ten separate coding threads were spun up from a single Codex conversation to build and test ten features in parallel, each starting from a clean context instead of one thread accumulating over 120k tokens of history.
  • A fresh agent thread reliably outperforms one already carrying heavy context, which is why starting new threads per feature is treated as a performance technique, not just an organizational habit.
  • OpenAI is sunsetting its standalone Atlas browser and folding that team's work into Codex's in-app browser, which now supports multiple simultaneous tabs.
  • The new Sites feature turns a single ChatGPT prompt into a hosted, shareable website, extending Codex from a coding tool into a full vibe-coding platform comparable to Replit or Lovable.
  • The hosts argue most software products today are effectively markdown files and configuration wrapped around existing models, so the differentiation now comes from what's built on top, not from rebuilding infrastructure from scratch.
  • Design judgment and precise written communication are named as the two remaining bottlenecks, since models can grade working code easily but still can't reliably judge good design or infer poorly explained intent.
  • The perceived AI education gap is described as bigger than the tool gap — most people already have access to agents capable of these workflows, they just don't know the techniques exist.
Takeaway

Four Codex habits that matter more than the model version bump.

WHAT TO LEARN

GPT-5.6 is a solid, cheap upgrade rather than a flagship-class leap, and the actual productivity gain in this episode comes from four repeatable habits: background computer-use QA, self-scoring loops, deliberate tool-stack choices, and running parallel agent threads.

02OpenAI Merging ChatGPT and Codex
  • OpenAI merged Codex into ChatGPT specifically to expose its ~983M-user base to agentic capability most of them don't know exists yet.
  • If you already use Codex for knowledge work, the new consumer 'Work' tab isn't meant to replace your workflow — it's aimed at the much larger group who still treat ChatGPT as just a chat window.
03Sol, Terra, Luna — the new models
  • GPT-5.6 Sol is aimed at developers and long-running tasks, while Terra and Luna quietly handle the ~70% of everyday chat tasks that don't need a large model.
  • A model's raw intelligence score doesn't predict agentic reliability — Gemini scored well but spiraled into redundant tool calls, while GPT-5.6 stayed token-efficient.
04The Replit/Lovable benchmark
  • The flagship comparison model still wins on the hardest, longest tasks in one shot, so it's worth deciding upfront which tasks go to a cheaper model versus the top-tier one.
  • A four-sentence prompt was enough for the flagship model to succeed on a complex build; GPT-5.6 needed a longer, more detailed prompt to reach the same result.
05Pricing of 5.6
  • GPT-5.6 holds 5.5's price while adding real capability, a signal that OpenAI is optimizing cost efficiency ahead of a bigger GPT-6 release.
  • Model subsidies won't last forever, so it's worth tracking token efficiency, not just sticker price, when picking a model for a given task.
06Computer use
  • Codex's computer-use runs in the background, so it can be handed an app to test end-to-end while the user keeps working in another window.
  • Assigning an agent 'test this app with computer use, fix what fails, repeat' turns it into an always-on QA engineer without writing a formal test suite.
07Record and replay for knowledge work
  • Record-and-replay converts a screen-recorded task into a reusable skill, which matters most for non-developers automating repetitive knowledge work.
  • The pitch is explicitly non-technical: record yourself doing a task once instead of writing a precise prompt or SOP for an agent to follow.
08Loops
  • An agent loop is just three parts — a clear prompt, a feedback set to compare against, and a stated success bar — after which the agent re-scores and revises itself.
  • Telling a coding agent to review and score its own changes, then keep fixing until it rates itself a 5/5, catches issues without standing up a separate review agent.
  • Scheduled automations (like a daily PR and security review) turn a one-off loop into a standing process that runs without being re-triggered manually.
09Loops for marketing
  • The same loop pattern works for non-code output: scrape competitor ads or reference thumbnails, then have the agent grade its drafts against that set until they match.
  • Design and creative work are harder to reinforce than code because 'good' is subjective — a curated reference set is what gives the agent something concrete to grade against.
10Plugins and social media scraping
  • Video-ingestion tools let an agent watch and score submitted content (like UGC ads) against top-performing examples, not just read a transcript of it.
  • Scraping any platform's top-performing content by handle turns competitor research into an input an agent can directly compare new drafts against.
11Tools
  • Matching the right existing tool to the job (a 3D engine's MCP, a video-capable model, a code-first database) matters more than defaulting to whatever stack the agent suggests first.
  • Asking the agent itself which tools fit a given job is a fast way to surface options you didn't know existed.
12Building block economy
  • Agents increasingly assemble pre-built services rather than coding infrastructure from scratch, so knowing which building blocks exist for your use case is now a real skill.
  • Humans are still needed to maintain and choose those building blocks — the agent stitches them together, but picking good ones is not yet fully automatable.
13Use multiple threads
  • Splitting work into separate agent threads per feature keeps each one on a small, focused context instead of degrading inside one long, cluttered conversation.
  • A fresh agent thread reliably outperforms one already carrying heavy context, so starting new threads is a performance habit, not just a tidiness habit.
  • An agent can spin up multiple sub-threads from a single conversation and later summarize all of them back into one document, so parallel work stays reviewable.
14New Sites feature
  • Prompting an app into existence and hosting it in one step (Sites) removes the separate deploy step that vibe-coding platforms used to require.
  • This is aimed at non-technical users first — the same capability is available to developers via a plain new chat, without needing the dedicated Sites tab.
15Updated browsers
  • OpenAI folding its standalone browser team into Codex's in-app browser signals that in-agent browsing, not a separate browser product, is the intended default.
  • Multi-tab support in the in-app browser removes one more reason to tab out to a regular browser mid-task.
16Agent-native operating system
  • Most software today is thin configuration wrapped around existing models, so the differentiator is what specific problem it solves, not the underlying model.
  • Design judgment and precise written communication are the two skills the hosts say still bottleneck output quality, since models can't yet self-grade design the way they can grade working code.
  • The bigger constraint right now is education, not tooling — most people already have access to agents capable of these workflows and simply don't know the techniques exist.
Glossary

Terms worth knowing.

GPT-5.6 (Sol / Terra / Luna)
OpenAI's model family announced with the ChatGPT-Codex merge. Sol is the flagship for long-running agentic tasks; Terra and Luna are smaller models aimed at everyday chat and knowledge-work tasks.
Codex
OpenAI's coding agent product, now merged into the main ChatGPT app as a 'Work' tab alongside chat, sitting next to computer-use and browser automation features.
Computer use
An agent capability that lets a model directly control a computer's mouse, keyboard, and browser to complete tasks, running in the background rather than taking over the user's screen.
Agent loop
A workflow where an agent is given a success criterion and a feedback set (reference examples or a working app to test against), then repeatedly revises and re-scores its own output until it clears the bar.
Record and replay
A Codex feature that converts a recorded screen session into a reusable agent skill, so a repetitive knowledge-work task only needs to be demonstrated once instead of scripted into a prompt.
Building-block economy
The idea that AI agents increasingly assemble pre-built services (databases, sandboxes, agent frameworks) rather than writing infrastructure from scratch, letting a single prompt stitch together tools like Lego pieces.
Agent-native app / mini app
Software designed to be operated primarily by an AI agent rather than a human, surfaced inside a chat interface only when the agent needs it.
Sites
A ChatGPT/Codex feature that turns a text prompt directly into a hosted, shareable website, without needing a separate deploy step to a platform like Vercel.
Resources

Things they pointed at.

02:20productGPT-5.6 (Sol, Terra, Luna)
23:37toolForeplay
27:44toolScrape Creators
24:49toolGemini (video ingestion)
29:00toolBlender MCP
30:45toolConvex
32:35toolEve (Vercel agent framework)
32:35toolDaytona
12:10toolLinear
00:45channelRas Mic's YouTube channel ↗
Quotables

Lines you could clip.

03:30
“It calls fifteen, twenty different tools. It, like, psyches itself. It itself experiences psychosis.”
vivid, quotable knock on Gemini's agentic tool-calling behavior→ TikTok hook↗ Tweet quote
27:01
“Ladies and gentlemen, your software factory is codex.”
tight, declarative punchline that caps the loops/automation section→ IG reel cold open↗ Tweet quote
35:18
“Ten threads were generated, Riley, and then I went to hang out with my wife and let Codex do the work.”
concrete, relatable proof point for the multi-threading workflow→ newsletter pull-quote↗ Tweet quote
44:55
“Nowadays I've realized most products are markdown files. Most products are markdown files.”
repeated-for-emphasis closing thesis on the building-block economy→ TikTok hook↗ Tweet quote
46:39
“I don't think there's that big of a tool gap. In fact, I think the tools are progressing too fast for people to actually adopt them.”
reframes the whole episode's stakes as an education problem, not a tooling problem→ newsletter pull-quote↗ Tweet quote
The Script

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.

metaphoranalogy
00:00OpenAI just combined Codex and ChatGPT into one super app, and they also released a new model that's nearly as good as Fable. They also made updates to their sites feature, which turns it into a full vibe coding platform, and they made updates to their in app browser.
00:15And so I have so many questions about all of this and how this actually works. So today, I'm bringing on Ross Mike, and we're going to break down all the new models, how the new platform works, and how he uses Codex as a developer, and how I use it as a marketer and founder. You're watching agent native.
00:32My name's Riley Brown. Let's dive in. Ross Mike, here we are.
00:37We're in the new ChatGPT app, which used to be Codex.
00:41OpenAI decided to merge ChatGPT and Codex, and also introduce three new models, one of which is very close to Fable five.
00:50Mhmm. I I wanna dive into a lot today, but I first wanna talk about the new platform. Do you wanna open up the work tab real quick?
00:59No. Right here. Everyone's favorite tab on Twitter right now.
01:03Yeah.
01:04Yeah. So, like, I'm just really curious.
01:06I I am still on the codex tab. I just don't understand the benefit for me. Should I just stay on the Codex tab?
01:12What do you think? On the Codex tab. It's not for you.
01:14The the Word tab, like Okay. The Word tab is not for us. Right?
01:16And, they are trying to get think about it. 983,000,000 users.
01:23That's more than the population of America twice. Add Canada in and it's still more. Right?
01:29So you have these people, and most of them are not using ChadGBT to the fullest. In fact, a lot of people are starting to think Claude is the agentic platform, and ChadGBT is just a place where I, you know, send texts and get questions asked. Like, I've heard a lot of people, uh, in my church writing, they'll come up to me and be like, yeah.
01:46Like, ChadGBT is cool, but, you know, with Claude, can do stuff. Right? So there's not only, like, a branding issue where people are just thinking of chat GBT as chat, but there's also an awareness issue.
01:58Right? This drop down specifically saying for getting work done is to tell people, oh, maybe I could do some knowledge work. Right?
02:05So if you've already been using Codex, you could stick to Codex even if you're doing knowledge work. The work tab isn't for me and you. I if I was an investor in OpenAI and I had some shares, I wish I had some shares.
02:16I would be bullish on this decision because there are millions of people who have no idea the full power and capabilities of Codex, and this is how they're gonna get used to it. 100%.
02:25Um, okay. So I wanna talk about the models a little bit. So have you so you've taken this model for a ride.
02:32You've tested the the different models that were released. Are your thoughts?
02:36Okay. Okay. What are your thoughts on actually, look.
02:40First, can we break down the three different models? So there was three models released. There was GPT 5.6 Soul, 5.6 Terra, and 5.6 Luna.
02:51Soul is the one that I think most people care about. I think naturally, they just care about the most powerful model. What are your thoughts what are your thoughts on Soul?
02:59So Terra and Luna are probably gonna live in the work tab. Right?
03:04This is probably what most normal people are gonna use. It's gonna live in the chat GVT tab because you don't really need super big models for almost 70% of tasks. But for most of us who are interested in these long running tasks, who maybe build applications with AI, SOL is the interesting one.
03:20Now, one thing about OpenAI and their model development is they're very token efficient and they're cheaper than the Anthropic models. Right?
03:29Um, price isn't necessarily what you just wanna look at. You wanna look at the token efficiencies. One of the first tests that I run is when I use a model, I check how many tool calls it does.
03:39And one thing about Gemini, um, our fallen foe, is Gemini, although on an intelligence index, is extremely smart when it comes to anything agentic.
03:50It just it just spirals into hell. It calls fifteen, twenty different tools. It, like, psyches itself.
03:56It itself experiences psychosis. Right? So the one thing that the GPT 5.5 model's really good at was being token efficient and was good at tool calling.
04:065.6 takes it at a different realm. Now, 5.6 in my opinion is not a fabled competitor.
04:13GPT six will be a fabled competitor. This is just a version bump. But even in the version bump, I genuinely think it is very, very comparable for what you get for your dollar, but not only that, in terms of sheer power.
04:26Right? Its ability to do really run long running tasks. One issue I had with 5.5, it was very needy.
04:33Right? Like, you give it a full plan, you tell it to execute on the plan, and it'll be like, oh, I did phase one. Should I continue on doing phase two?
04:41It's like, I gave you a whole freaking plan. Do the entire thing. Right?
04:44Come back to me once it's it's done. 5.6 is a lot more eager, is a lot more faster, um, and it's token efficient as well.
04:53But it's not a Fable equivalent. Like, I I I definitely think the Fable equivalent will be a six. Right?
05:00And this is why they did a version bump of 5.6 versus
05:03GPT six. But all in all, incredible model. Yeah.
05:07GPT 5.6 passed my I don't know if you saw this. My I have a new replet slash lovable benchmark. Yes.
05:14Yes. Can these can these models in one prompt create a lovable clone, which is a mobile app like a Swift iOS app that builds web apps just like lovable and replet?
05:27When I asked Fable to do it, my prompt was like four sentences. I it's in my video. It's pinned on my Twitter if you guys wanna see my prompt.
05:34But I basically said, I want you to create a Lovable clone. You use sand use Daytona for sandboxes, Convex for database, and I think I use Anthropix models to generate the app.
05:47Mhmm. And in one shot, Fable didn't hesitate, and it just did it in four prompts.
05:55I will say 5.6 required me to create a longer prompt because I had to, like, give it more instructions.
06:02And that makes sense. Right? Because, again, people are comparing 5.6 to Fable five.
06:07I don't think it's fair. I think the mythos class models are in a different class. If you notice, every, like, four, five, six, like, there's always been, like, something like it's been a monumental shift.
06:19The only, like, point shift that was monumental was opus four five. I think opus four five is when we all realize, like, woah. Like like, this is not a joke.
06:27Right? So I don't think it's fully fair to do, a one to one comparison between 5.6 and Fable because Fable is just a monster. And I genuinely think with the considering cost, it's because important thing to consider cost.
06:40We're not gonna get subsidized forever. Considering cost, I think it's very wise for people to think and start to aggregate. Okay.
06:47What tasks can I chat can I trust a model like 5.6 with? What task do I give Fable with?
06:53Right? If you're like, I don't know, changing the color of a landing page, uh, maybe you don't use Fable five max. You know what I mean?
07:00Maybe we throttled it down to low. Maybe we even give it to Opus 4.8. Right?
07:04So Fable is an incredible model. Like, some of the things, Riley, I built is I built a Daytona alternative.
07:10Right? So spinning up, like, your own sandboxes on a bare metal server. Um, I built a game plane, which is this dashboard where you can deploy your own, like, game servers like Vice City, Minecraft, and, like, it's modded Vice City.
07:26Like, I have like, it's not fully functional, but I'm trying to build, a multiplayer Vice City. Fable's doing it. Like, some of the stuff I wouldn't be able to do.
07:34Eve, which is Vercel's agent, uh, framework, I ported that so it can fully run on Convex. Right?
07:41This is something I could have done, but it was taking me three or four months. And this is like three or four months of like, this is what I'm focused on. This is what I'm thinking about.
07:51This is all I'm grinding on. It took it eight hours. Eight hours, one prompt.
07:55It did the entire thing. All tests passed, fully deployed on Vercel, fully working chat app. Right?
08:01So Fable is in a class of its own, and I think the comparable model that they're probably cooking up right now is GPT six. 5.6 is a class above Opus four eight, a class above five five, much better than these models.
08:18But in terms of sheer intelligence and getting very hard tasks done, nothing touches fable.
08:24Yeah. And it's it's also the same price as 5.5 too. So it's it is incredible, by the way.
08:29I I don't think people, like, grasp that. Right? Like, it it is difficult to build a very smart model, so kudos to Anthropic.
08:37But it's very difficult to get price, right, to make it efficient. And I think OpenAI is building that foundation of building efficiency, and then they're just gonna train the super duper new class GPT six model.
08:50Right? So I I think cost is something people should really start thinking about, um, we're drinking the Kool Aid too much because one day the subsidies will end.
08:58Yeah. I I agree. I agree.
09:01What I do wanna do is I I really wanna selfishly, I want to kind of get some of your workflows. I wanna understand how you you how you've been using Codex, your workflow from a tactical perspective,
09:12and some of your favorite skills that you've created. Yeah. I you know me.
09:16I get it to Excalidraw. So there's a couple things that OpenAI, ChatGPT, Codex, whatever, these names are all over the place that do phenomenally well.
09:26Right? The first one, and I beg people, just use this, is computer use. This model is incredible, incredible at using your computer.
09:37Right? So much so that I made the decision to get a mini a mini PC so that I can remote view into this and I could tell it to use this computer. It's that good.
09:49Now you might ask, what are you using computer use for? To fill out forms, to do this, that, and the third. In terms of application development, my number one favorite thing to do right now is test my app.
10:00As simple as that. You don't even need a skill for this because the model and the agent are so good. So let's say I've developed
10:08automatically now. Like, I noticed don't even have to do testing.
10:11It I I come back thirty minutes later. I'm like, why is it taking so long?
10:16And I see that it's, like, testing it, like, over and over and over again, finding edge cases, and just doing it. Anyway, sorry to interrupt, but it is the computer use and and the browser use it has is genuinely something that you cannot
10:29we could not explain it properly. You must experience it. It is incredible.
10:34Anyway, keep going. Yeah. So, like, uh, you know, there have been talks about loops and stuff.
10:38One of my favorite loops that I've been using, and you can use this with really any agent that has computer use, but I find that OpenAI Codex is the best at this because their computer use is just great, is as follows. I'll connect something like like linear.
10:53Like, I use linear for, like, issue tracking. You can use whatever Notion, whatever whatever. Right?
10:58And I will basically tell my agent, you'll build this feature, copy paste, or you can even connect the MCP tool and attach it as an agent. And this is the loop I run.
11:08I literally tell the agent, hey, build this feature.
11:13The success criteria for this feature is you use computer use. If it does not succeed, fix, repeat, and once it's done, come back to me.
11:22Right? Um, this is me, by the way. I just asked now you might wonder why that popped up.
11:27I just asked ChadGBT to run this app locally, and then I'm gonna say, hey.
11:33Can you test this app using computer use end to end? And I'm gonna hit enter. So I'll get it to use computer use as an evaluation step.
11:44Right? Because if I'm building a feature and that feature is somewhat UI centered, right, it's a form submission or anything like that, Code isn't just enough.
11:52Right? I need my agent to QA test this. And even in like a a a a a normal computer engineering team, there are people who are going to you know, there are QA people who test this.
12:02Right? At my very old corporate job, Riley, I used to build a feature, and there was one person's job. Their whole job was to break what I built.
12:10Right? I can now do that with Codex. And now you're gonna see Codex itself test this app out.
12:18It's just testing it up. It can spin up the browser, and it's just testing it. Yeah.
12:23So it opened it up. It says the browser pass succeeded. It basically made sure that it worked on browser.
12:28So it's just going. Right? And the cool thing about Codex's computer use is I can do my work while it does it.
12:36Right? With with Anthropix computer use, it actually takes over your mouse and your screen Yes. And you kinda have to let it work.
12:42Right? So This is very important.
12:44This is very important. When you use computer use sorry to echo what you're saying because I think it's really important. When you use computer use when you use computer use on codecs, they can do it in the background.
12:55When you use it on Claude code, you'll notice there's just like kind of like brown gradient that takes over your computer, and you're just like, oh, I have to wait for this agent to be done using my computer. But Codex, you'll I'll switch to a new app tab, and I'll see that it's working in the background, which is which is a mind blowing experience, which makes me wanna get the most powerful computer ever.
13:19And and I wanna just I wanna have a 100 computers with agents working on them. Like, it's crazy. And this is the thing.
13:25Right? Like, computer use and and I might pivot on a point here. Like, you really need to start thinking bigger.
13:32Right? If you have an agent that can use your computer well, my question for you is,
13:37what are the things that you can now do? Right? What are like, let's say there's, a spreadsheet with thousands of rows and there's certain things that even an MCP connector couldn't do.
13:47You can now do that with computer use. Let's say you want your app fully tested end to end. You can now do that.
13:55There's so much you can do with computer use. Right now, I have two browser instances. I don't know if you could see my screen.
14:01And maybe this is the best product to pick to pick, but it spun up two different browser instances, and it's basically running commands on this virtual terminal I have. Right?
14:11And look. It says launched and left a visible machine back end.
14:15So this is stuff that, again, these things were people's jobs.
14:21Right? I remember at my old corporate job, Riley, I had someone whose sole job again, I build a feature, they run the feature on their on their computer, and they would spam and break it. Now one thing that I do is I'll ask the agent to create a test plan for my app because it has context of how my app is and test it using computer use.
14:40So it's going to use the app. It's going to identify, you know, what features work and what don't.
14:46And guess what I could tell it to do? I can tell it to fix my features. So I would say homework for everyone watching, at least at least, please, me, at least thirty minutes.
14:56Just just spam computer use. Tell it to do the most random things. If you have an app being built, tell it to test out your app using computer use.
15:03This is, I would say, the biggest Can I up for me in terms of workflow? Can I add one thing here?
15:09The best way to use computer use. If you are not a developer and if you have no interest in vibe coding, let's say you're a marketer, and and I honestly think I use this way more for marketing tasks than I do coding tasks now. Um, but OpenAI released a new feature.
15:24It's called record and replay. What this does is it allows you to tell the agent to record your screen so that you can just use your computer for up to thirty minutes, and then it will take that recording and convert it into a skill so that the agent can do it later. I 100% think that this is the future of of computer use for knowledge work tasks.
15:47It's not gonna be writing this perfect prompt and like writing out an SOP for an AI agent. It's simply going to be recording your screen. And a lot of these tasks are repetitive and annoying.
15:59Like, there's so many tasks that take me ten minutes manually on a computer that if I could delegate to an agent, I would so much happier. But so, like, the full process is literally, like, uh, I want you to record my screen and turn what I do into a skill. Like, literally, that's all you have to do.
16:16I digress. Let's get back to the list here. Number two is loops.
16:20Now
16:21I know this is a lot of controversy behind loops, but hear me out. I think their loops make sense in certain card categories when you're building your application.
16:32And because codecs mixed with g v d 5.6 sole, super powerful model, one thing that you can do is you could have an automated, uh, review cycle.
16:43Right? So every time, like, you push changes or you make an update, you can tell the, uh, the GPT 5.6 SOL model to review the code and to assess it and give you a score from, like, one to five or one to 10.
16:59Now I use an external code review agent, but you don't even have to do that. You could tell SOL itself to give a review of the code it changed. I would do this in a new thread, and then give it a review.
17:09One out of five. Most of the time, you'd be surprised the agent that wrote the code itself will review itself pretty harshly and it will tell you, oh, this is a four out of five or this is a three out of five. Sometimes it's a five out of five.
17:20And then what I would do is I would tell the agent, keep fixing until you give yourself a five out of five.
17:31Another type of loop. I want I know I'm jumping ahead, but these two tie together is using our automations. So if you go on the cursor tab right here, I think they've changed the schedule, but codecs tab.
17:43But okay. Stop. Sorry.
17:44It's okay. I I I I did wrong. I did wrong.
17:47Um, if you go to the codex tab, I think they changed it to schedules now. I can, um, I can basically say, at 8AM every day, I want you to review all my open PRs, and I want you to give me feedback on which ones I should close, which ones I should keep open, which ones, uh, need further review.
18:09Um, I want you to do a thorough security review of every single PR and give me a report. And if you find there's something necessary to fix on the spot, spin up a thread and fix it.
18:21So just something like this. Right? Mind you, this with computer use.
18:29You know, people talk about software factories. Ladies and gentlemen, your software factory is codex. The only thing that was missing is a super powerful model like GPT 5.6 Soul.
18:41So you can write an automation like this, works every single morning, specific time. It will review all your open PRs.
18:49It will review all security issues that you have, and you just have a report. So, you know, we talk about being managers and managing software factories. That really wasn't possible until models like 5.6 and Fable came about.
19:02What really is the issue now is one's imagination. And I think what we what I would like for people to watch from this is, like, how far can I push this? Like, what's the weirdest thing that I can build?
19:15What does my heart desire? You know what my heart desire? I wanted to build a GT and let me show you some screenshots.
19:22I wanted to build a multiplayer GTA Vice City game because they don't exist, and I like GTA Vice City, the classic one. And I'm going back and forth with this. It forked a Vice City fork that runs on browser, and then I asked it, hey, I want multiplayer.
19:39It went about four hours, and guess what it did? I have multiplayer.
19:43The one screen is the gentleman standing. The other screen is the person in the car. Now there were some bugs.
19:50We had to go back and forth, but can you imagine, Riley, as someone who runs a team, can you imagine how many engineers that would have taken? How much time that would have taken? Right?
19:58So all these tools that you have access to, the computers and all this stuff, what this has now allowed you and me is to build bigger and wilder things. So going back, because I know I go up on tangents Can I can I Can we go back to this? Can I add something here?
20:13I I wanna make this tangible for non coding tasks. Can you go to the Excalidraw real quick? Yes.
20:18Just just talking about loops here. So, basically, what he did is he
20:24basically told the agent to review its work, give it a score, and it basically decide like, if it doesn't meet its own certain, uh, doesn't meet some certain criteria, it loops back and it starts again until it reaches a certain score.
20:40That's basically the essence of a loop. And this can be applied to anything in and so, like, obviously, I'm a marketer.
20:47Right? I'm a founder Actually, I would love to know. Yeah.
20:49How do you use loops in marketing, Riley? Because, like, I This is the way we do it. Yeah.
20:53So there's there's a let's let's take ads for example. So there's an there's a an API called Foreplay.
21:01Very weird name for a company, but it's called Foreplay. And it allows you to scrape ads from any company and you can basically pull the videos, pull the transcripts from all of the top performing ads, and you can sort them by longest running.
21:17And if a company is running their ads for the longest, you know that that one probably converting the most. Right? You can you can't tell how well they're converting or what the ROI is, but if they're running an ad for a really long time, you know that it's a it's a good ad.
21:30And so what you can do is is you can basically in marketing, is you can pull all of your competitors really good ads or really well performing ads.
21:40And if you're asking AI to create 20 ad scripts.
21:45Right? 20 ad scripts for your company. Right?
21:47And you're you can give Codex all of the context of your company and it writes these 20 scripts. It can basically compare the scripts that it wrote to the high performing ads and give each one a score.
21:59And you can have Codex go in a loop until all of them are like a certain criteria or comparable to the other ads. That's just one example.
22:09Another example is YouTube thumbnails. And the only way to do this in marketing right?
22:15Because if I don't know if you would agree with this, but coding is is a lot more it's a lot easier to say good job or bad job. You know?
22:22And that's why design that's why that's why AI models are bad at design. It's because it's ambiguous as to whether it is a good job or a bad job.
22:30Whereas, you know, design, scripting that's why that's why AI is not good at script writing is because these companies that are training it, it is ambiguous whether it's a good job or a bad job.
22:41How do you reinforce an AI to do a good job if the diff the people doing reinforcement on it all have differing opinions. No one knows what's good.
22:49Where coding, it's very easy to tell. And so the only way to do that is to gather really high quality examples and give it to an AI and then tell it to evaluate its work in comparison to all of those examples.
23:01Right? If you give Codex 20 different thumbnails that you like, it has a really good sense of what you like and it can grade the thumbnails that it creates according to the thumbnails that you like.
23:13And you could just say, like, keep going until the thumbnails that you generate are similar to those 20. Does that make sense? And so that's kind of like how we're thinking about it.
23:23And with one tool that I have Codex is it can use the Gemini video ingestion feature.
23:30Right? Gemini, you can upload a video to Gemini and it can actually like watch the video.
23:35It can pull subtitles from it. It can see all the frames. And so Mhmm.
23:39Codex has a tool which uses Gemini and it can analyze my video. So we have UGC creators and and other creators for our company will submit videos and I'll just have Codex analyze them and it will compare those videos that they submit to the top performing videos that are run by other companies, and it'll judge them.
24:01And so these are kind of the loops that we're starting to create is comparing stuff that we do as humans in our company and then also stuff that the AI generates and compare it to really high quality examples. And it basically loops and edits them until it reaches some quality bar.
24:17Does does that make sense?
24:19Yeah. And, like, to further, like, just solidify what you just said. A loop is as simple as follows.
24:26You have a well defined prompt. Right? In this case, for Riley was, you know, I want a YouTube thumbnail.
24:32You have a feedback engine. Right? In his case, it was 20 other thumbnails that were really good, and he wanted these thumbnails to be similar.
24:41And like a developer's case or you were building an application case, it's telling the agent to use computer use. And honestly, nowadays, you don't even have to tell it. It will use computer use.
24:51Use the app, and if it functions just as well, then it will that's that's basically your feedback engine. And then you have a well defined success criteria.
25:01And you're basically telling the agent, until we reach here, using the feedback engine, keep on going. Right?
25:07So loops apply to anything, any industry, any space.
25:12You just need to know what your what it is that you want. Most people don't know what they want. Most people don't even know what they want for lunch today.
25:19You need to know what success looks like because the agent's not gonna tell you what success looks like. The agent doesn't know what success is. And then you're gonna give it a mechanism for it to give feedback.
25:29Right? In Riley's case, was he had top performing UGC videos that were used as comparison.
25:35In, uh, building case, it would be using computer use or yours you've specified that you want the features to do x, y, and z. And this basically, the agent itself will keep on looping.
25:46You don't even have to tell it to loop. It will keep on going until it reaches what is defined as a success criteria. Right?
25:54And this is how you use something like Codex in any single field, in any industry with the tools that it has, with the model that it has, and this capability, there's honestly nothing we can achieve nowadays.
26:06100%. And and I think that's why plug ins are so important. You know, like, and I talked about the ads plug in.
26:13There's also another API. It's called Scrape Creators. And Scrape Creators allows you to literally download any video from any platform.
26:23So I could say,
26:24I want you to look up I don't even know that. What's it called? Scrape Create
26:28Scrape Creators. You can look up you could say to any agent.
26:32You could be like, hey. I want you to go to Greg Eisenberg's Instagram and download his top 10 videos.
26:37Right? And it will just scrape them. And because of my Gemini skill, right, which which allows codecs to gather as much information from the video as possible, because to my knowledge, Gemini is the only one where you can actually video you can upload a video to the model.
26:52Think any other model can do that. Yeah. Which is so the most underrate that's, the only thing I would use Gemini for.
26:58They're kinda behind on everything else. And I don't mean to trash them, dude. I Faith Logan Kilpatrick, if you're for some reason watching this, which you're not, I I love you.
27:05Keep going. But yeah.
27:09That like, if you give the agent the ability to go find high quality examples and then you just tell it, like, hey. I want I want you to create video.
27:17Like, make sure this is as good as Greg Eisenberg's video. Make sure that this is good as RPN's videos on Instagram. It'll just go to Instagram, download his videos, watch them, understand what good looks like, analyze the script, and then it will make sure my videos are edited in the same way.
27:36And that's kinda how I and then even this is really useful for manual processes as well. So, yeah, that's just kinda how I've been using it for marketing.
27:45Yeah. So loops, again, well defined prompt, feedback, uh, engine, and then a well defined success criteria.
27:52This again, I didn't know he was doing that for marketing. I just learned something new. I've been doing all this stuff manually.
27:57Right? I believe me, I should've had a loop going on. Right?
28:00So this literally applies to anything. It's just you just have to have a mindset shift and start really to think how far can I push these models and codecs?
28:09And I think that helps a lot. Yeah. Um, so this list.
28:13So we had computer use, loops, anything else?
28:16Number three, um, and you kinda said this, and I wanna be careful on what I mentioned here. But number three is tools. Right?
28:24Like and when I mean tools, like, from a development standpoint is what stack are you using? I think, um, you know, back in the day, back in the good old engineering days, um, we would very much take pride in building a lot of the stuff and owning the stuff.
28:41Now it's gone to the point where speed really matters, and we don't really care for that. Right? So I wanna make sure I'm using tools and things of that nature that really, really help me develop whatever software I wanna develop.
28:55Right? Using the right tool for the right job. Right?
28:57If you notice, the agents for the most part default to react when you tell it build me a web app. Right?
29:03With login, this, this, that, and the third. I think having sort of, like, that sort of knowledge and knowing what it is that you need to build the thing that you're trying to build is necessary. If I know it builds three d stuff, the first thing that comes to mind is Blender.
29:15Blender has an MCP. I go to Blender site, copy the MCP instructions, and then boom, I paste it.
29:21Riley wanted to do video ingestion. I didn't even know, like, a model could do that.
29:26Right? So he goes to Gemini. Right?
29:28Using Gemini. I'm only gonna mention this because you said this, and I I I will preface from the beginning.
29:36I do work there, but seriously, seriously, seriously, Convex when developing applications, and I know people are gonna call me a shill, Riley, but, uh, sincerely, there isn't an app I don't build without Convex, and this one reason is the only thing. It's entirely code. And I'm getting to the point where I want the agent to do more of the work, more of the, like, like, the the the the tasks, and I really just wanna be the manager who kicks their feet up and makes decisions.
30:05Right? So what tools can I give my agent that gives it full access, that gives it a large surface area of things to do, like that gives a full power, but that allows me to just make decisions so that it I just say yes or no?
30:19Right? So the tool stack that you use is important to this to the stuff that you wanna work. So when you combine computer use with loops with the right tools for the job, I don't think there's an app that people can't build.
30:32I think the one thing that blocks a lot of people is not using the right tool for the right job. And what's interesting is if you ask Codex what tools are best for you,
30:41Codex itself will tell you, okay. This works for me. This works for me, and these are the reasons why.
30:45And I think this gets at kind of what, um, I think Guillermo Rausch from Vercel talks about a lot. He talked about this on the Naval podcast that there's this whole building block economy building around agents where agents don't have to build any, like, a lot of things from scratch. Companies are building these building blocks that AI agents can just use.
31:04And if you know what you want to use like, you brought up Eve earlier, which is Vercel's, like, the agent framework, and you can just spin up an Right? And the way that I and I I was using their AI SDK, I think is what it's called Yeah.
31:19For my replet clone. That's what and so when I asked Codex to build replet, it used it used Convex for the database, which so it just used this external tool, and then it just used the AI SDK from Vercel to have an AI that's the agent that generates the code inside the app.
31:42And it just used these and then it used Daytona for the sandbox because you need a sandbox to generate and run the the code for every single app. And that's basically like, all the agent did was kind of connect these prebuilt
31:57building blocks like Legos, and that's how it creates all these different things. And and one thing I I I will say that most of these tools, like you mentioned Convex, Eve, what's interesting is, and especially with Eve, I really like Eve, and the reason being is the decision behind how an agent is constructed.
32:15Now this might seem very high level, but it's so simple. If you notice Riley, you the the file names, the folder names kinda give you a description as to what it does.
32:26Agent dot t s. So maybe this is where the agents define. Skills folder.
32:30Oh, this is where the skills are gonna live. Tools. This is where the tools are gonna live.
32:33The sandbox. This is where the sandbox is gonna live. Channels.
32:36This is where, you know, the agents communications channel are gonna be configured, connections. So so they constructed the framework with file names.
32:45And the reason why they did that is the agent, guess what, has access to all the files. So when it comes to making changes and reproducing things, it just makes it so simple. Right?
32:55And that's why, like, build like, using tools that really care about how your agent uses, it is very important.
33:03To Guillermo's point, there is going to be a whole economy where some people's companies are literally plug ins, markdowns, and MCPs.
33:11Right? And the focus is how good can the agent use it? Can I give it one prompt, and can it build everything from scratch?
33:18And the reason why the Riley benchmark passes with replet, the new models pass with replet is because the models know of these tools, and instead of building the thing from scratch or using AWS or using whatever low level primitive, they're just going to build using these tools, and it just makes it so much easier. And this is why we still need developers.
33:37I still want developers, really smart people to maintain those things. Right? I don't think we've got to the point where AI could fully maintain those things, but that that's an important point to realize that we I wanna give my agent a tool that it can fully use and maximize on.
33:51And sometimes this comes through experience. So I would be someone who uses a ton of tools and see what you like. About that.
33:58Yes. 100%.
34:00Was there, uh, another was there number four, or was that the the finality of the list? Well, number four is this. Number four is and and this is gonna be controversial.
34:10Use multiple threats. I think there is this, um, like, instinctively, you have this idea where, like, you're in a flow state, you're going back to back with the agent, you know, you're there's this notion of sticking to one thread.
34:26May I present that using multiple threads does two things. First and foremost, you get used to this idea of compartmentalizing features that don't touch each other or work that doesn't affect each other so that you can work on multiple things at the same time.
34:45But number two, what it's important to realize is, uh, different threads is a fresh pair as as as fresh context. So now I'm basically using the agent at its most capable state.
34:57If I got a fresh agent versus an agent with a 120 k tokens context already taken, the fresh agent will outperform any single day of the week.
35:07And what's cool with Codex, Riley, I don't know if most people know this, you can spin up multiple threads from a single thread. Now, um, I I I did this on a on a work codex account, so, unfortunately, I can't show everybody. But one thing that I did was I built a a full thing at work, and then I told in the thread, spin up as many threads as they are features.
35:27Use computer, to test each one, let me know if everything passes. 10 threads were generated, Riley, and then I went to hang out with my wife and let Codex do the work.
35:38Right? So it's not is important.
35:41Tool to my knowledge. When you're in Codex, you can ask Codex directly.
35:45You can say, hey. I want you you could literally just list all of the different things. I want you to do these 10 tasks in 10 different threads, and it will spin up a new session.
35:56You'll see them being spun up in the side, and it's incredibly interesting. One other thing that you can do with the multiple threads is you can say to Codex, hey, I want you to please summarize all of the threads that we have today and create a little document of all the things that we've created.
36:12It'll just read them. So it so by default, yeah, you have these isolated threads, which is nice. But if you ever want context from other threads, you can just say, hey.
36:21Earlier today, we were working on these different things. And since I do a lot of knowledge work and I'm constantly doing my email, you know, you know, you use it as a builder, and that's great.
36:29Right? You're a builder at heart. You know, I'm a marketer.
36:32I'm kinda operate a company, like, from the marketing perspective. So I'm just in all these different softwares. I'm constantly emailing different people, um, writing different scripts.
36:41It's really useful to just get an AI to just summarize all the different things I've done. And it'll just create a concise one pager. And you can even ask it to include little links to the thread in your little document that it creates because each session has its own thread.
36:56Be like, you created this here. And if you click on it, it'll take you to that thread directly which is really a really cool a really cool feature in my opinion.
37:05Can you open up Codecs again? I wanna discuss three things that ChatGPT talked about, and this is their new sites.
37:14There it is. There is new sites. Sites on the sites.
37:16Oh, right there. Okay. Okay.
37:17Okay. There you go. There you learn.
37:19Nice. So okay. So this is a new feature that OpenAI announced,
37:25which
37:26they announced this a few weeks ago, but it was only available for Teams. Right? So notice how when you clicked on that, now it says create a website that and just say that is Hello World.
37:34Let's just do one that's really quick, um, and then make it like the fastest model that we can do so we can show it. Yeah. Oh, I love this by the way.
37:42Uh, although Yeah. It's lovely. Very lovely.
37:44Yes. So you can fire this off. And what this will do is this will create a hosted site that you can send to people.
37:51Right? This is gonna be hosted on the Internet, and you can share this with people. And I believe that as part of their announcement yesterday, this is available to everyone.
37:59This was only a Teams feature. Now I think you can host sites and share them with friends, um, like you can do on something like Replit or Lovable.
38:07Right?
38:08Notice how you had to you you go to sites. There's one thing I noticed, and I click create create new site, and it's a prompt with, like I'm I'm assuming this is, some sort of, like, combination of skills and MCP.
38:19Yeah. So design feature app. So the reason why I bring this up is this is how capable, like, this is how, like, a capable model is important.
38:28A lot of this stuff, a lot of the cool updates they made were product facing. It's still the same underlying model, and it just shows you that Soul in particular covers a large surface area.
38:38And I go back to the first point I made in our previous part, The best coding model
38:42is the best everything model. Yeah. I didn't even do it correctly.
38:45No. 100%. You don't even need to go to the sites tab too.
38:48Like, you press command n on on codecs. Let me You can press command n and you can just wait.
38:56Just, yeah, get rid of that, I guess. No. That that that's what happened when I pressed command n.
39:00Yeah. Yeah. Yeah.
39:01It took you to a it was already written. Let me do it. So, like, you just go to a new chat.
39:05Command n? Yeah. And then you just type at sites.
39:09And there you go. You type at sites and then you could just build a site. You can vibe code it.
39:13And you can and, like, you can host it on their platform. And then once you like it, you could host it on Vercel if you're used to using Vercel. But this is an easier way for people who aren't technical to vibe code.
39:24Like, they've basically turned Codex into a full vibe coding platform with this. And it, like, it allows you to iterate. It gives the model an ability to use this plugin.
39:33And I think if you were to ask for a website, it would just automatically host it in this. You should. And then you can open this oh, whatever the link that it creates, like, you can open it in the browser, which is really cool.
39:43That's awesome. Yeah. Another thing I wanna bring up, while it's loading, actually, can you open up the side panel and hit the thing on the top right?
39:50On the right side. On the right side. Yes.
39:52Here. Yeah. So you can open up a browser.
39:56Just open up the browser really good. Really Yes. Really good.
40:00So as of this latest update, now it can open up multiple tabs. So, like, I I was using it know that. Yes.
40:08Can open up multiple browser tabs. So you can open up new, uh, multiple browser tabs now. Um, and, yeah, it's a full browser.
40:18So one of the announcements they made yesterday is they're sunsetting Atlas. So OpenAI is sunsetting Atlas, which they used to have a team working on Atlas, is the browser. And they've cut or at least merged that team.
40:29They basically sunset this private app, they're putting all of their effort in making the in app browser really good inside Codex. And so that is what makes me so bullish on AI agent native sites. It's like so many people who use ChatGPT are now gonna have this tool.
40:46And, like, when they ask for something, instead of it opening up Chrome and in some side browser, it's just gonna open it up in the browser right there. And that's You used a term for this. You were using a term for this.
40:55What was it like mega app or super app? Super app. Right?
40:58So sue a super app is is this is Codex. Right? Which you can think of as like an AI operating system.
41:06An agent native app or a mini app is an app that is meant to be used by an agent first. Like, I think it's like agent first. Dan Shipper was talking about this on this last he's been building agent native apps.
41:18These are apps that he never goes to unless he's using an agent. An agent can basically pull up this little email mini app that shows him the app that the AI drafted and he can just send them off or, like, provide feedback to Codex.
41:32And so these are these apps that I think are going to emerge, which are just apps that are made to be used by agents with you.
41:40Like, your agent can show you things, but, like, there's buttons in it, but it's meant to be agent native. And so I think that might be what you were asking about. The way that I see it, it is a it is an operating system.
41:50We're seeing the early early glimpses of what an AI operating system if you were to buy a computer and it just had this loaded up, it could you could basically do everything. And you just chat with an AI, a thing happens, or an interface loads up, you use the interface, and then you move to a different chat, uh, interface.
42:08And so, yeah, that's what makes this so exciting for me is, like, we're literally watching a kind of it's just a new way of using a computer. We're just in the early days of it. So it feels kinda like ChatGPT, but, like, if you get good enough at using tools like this, I'm not even just saying use Codex, even though I do think it's, like, the best in this paradigm.
42:28You'll you'll really start to become more agent native, and you'll be like, wow. This can literally do anything. Anyway, we covered a lot today, Ross, Mike.
42:37I guess my closing question to you is this. What are you excited about in terms of codex? And, like, what are the next few things you wanna work on?
42:44And and, like, in order of, like in terms of acquiring skills around codex, what do you what do you wanna do next?
42:52Yeah. I think, like, it it goes back to like, and I tweeted this. It goes back to what I said earlier.
42:56Like, I'm really starting to think bigger. Right? I'm really, like, starting to, like, those, like, dream ideas, apps, or businesses that I had in my mind that are like, man, maybe I don't have the intelligence level or I don't have the time or I maybe I just don't have the wit to figure out all the, like, know, nitty gritty stuff.
43:12I really can with, like, you know, with the latest and greatest model and with a harness like Codex. Right? So the big thing for me right now, Riley, is I'm like writing this running list of, like, things I've wanted to build whether it was a kid or whether I was now.
43:26And I'm just building stuff. Like, I have, like, eight to nine different threads that are running, that are building things. So to me, like, my eyes have widened mostly because of Fable TBH, but then using GPT 5.6 Soul because we're gonna lose Fable in a bit.
43:42It made me realize that, oh, I have a model that's really, really good for a fraction of the cost that's able, needs a little bit of guidance, but that's still able. So me, my mind has expanded. Now in terms of, like, things that I need to learn and skills I'm trying to acquire, I think it's the things the models are not good at.
44:00Right? So design. I'm really starting, like, I have a a running catalog of, like, really good design.
44:06Right? So design inspo. I'm trying to train myself on, like, okay, colors, typography, spacing.
44:12Because believe it or not, even as good as the cloud models are with design, at at a certain point, we'll start to realize what are the AI sites and what are not. Right?
44:20So design is a big one. And the other thing is actually writing, like, and being able to explain stuff.
44:27Like, as the model gets better, like, sure, it'll have intuition on what you mean, but imagine you and I had the ability to communicate effectively, explain exactly what it is we want, what it is what we expect, and what the feedback loop is. I find myself that the better I'm getting at explaining these things, the better outputs I'm getting.
44:44So design and writing is a big thing that I'm trying to improve myself with, but in terms of what I'm doing, I'm really just trying to push even further. What can I build? What can I do?
44:54And one final thing I'll say is, nowadays I've realized most products are markdown files. Most products are markdown files.
45:02So don't spend time building a markdown file. Spend time building something that is actually useful, um, and important.
45:10And if those things are difficult, well, now you have access to the latest, greatest model. So that's that's pretty much my thoughts around what's going on right now.
45:19Yeah. And and and for me, I'm just gonna be honing my marketing workflows. I'm gonna spend a lot of time on this.
45:25I think this is the most under under explored things, like using it for operations and marketing and basically using it as, like, my operating system. And
45:34I think you're one of the few people who who got jumped in with the engineers and learned how to use the tools with them, but now you have this unexplored field that a lot of people in this space are not really privileged to these tools. And even if they are, they're not really that good because I'll see the content on Instagram and whatever, and sometimes I cringe at the advice that people, like, you know, give.
45:55Like, I remember I saw this one video where the person was like, oh, like, you can use, uh, GLM four point whatever, and you could save a lot of money.
46:04It's such a great model. And I was like, no. It's not a great model.
46:06So education
46:08the education gap is massive. And I I think that's one thing I wanted to bring up earlier, uh, which is you can build all these different products. You can do so many things, and you, uh, I honestly think the biggest opportunity is education.
46:22I think there's a huge education gap in the market right now. I don't think there's that big of a tool gap. In fact, I think the tools are progressing too fast for people to actually adopt them.
46:32And so that's why I'm focusing more on content now is because I've I think edge people are not properly educated, and people want to be educated. There's just a massive amount of terrible content out there, especially on short form platforms.
46:46And so if you can figure out a way to be a voice of reason, teach people how to use these tools, and then even go into businesses and just help them out by, like, by implementing these tools and teaching them the workflows, you can make so much money, especially as these tools get better at general knowledge work tasks.
47:02Actually, it's already good enough. It's just about how do you actually, like, conceptualize the operating system and configure loops but for marketing tasks or configure automations but for marketing tasks or configure the tools that we brought up, plug ins.
47:17What plug ins and tools do you need to use? And and and, like, what is the workflow? And so that's what I'm really focused on right now.
47:24Agenic workflows for noncoding tasks and how to conceptualize it and teach people at scale, that's just kind of, like, my main focus. So make sure you subscribe to the Riley Brown channel.
47:36Yeah. Let's go. Anyway, it's always a pleasure to have you on the pod.
47:39Um, I'll hope to have you on soon. Um, we should do this, like, every few months, dude. I love it.
47:44I'm I'm game. Again, thank you, Riley, for having me. I really appreciate the opportunity, and thank you everybody for watching.
47:49I hope you learned something.
The Hook

The bait, then the rug-pull.

OpenAI folded its coding agent into the main ChatGPT app and shipped three new models in the same breath — Riley Brown pulls in developer Ras Mic to figure out what's actually new versus what's just repackaged, and to hand over the four workflows they lean on every day.

Frameworks

Named ideas worth stealing.

16:06list

The four things Codex does well

  1. Computer Use
  2. Loops
  3. Tools / Stack
  4. Use multiple threads

Ras Mic's running whiteboard list (built live in Excalidraw across the episode) of the four Codex capabilities he leans on: background computer-use for QA, self-scoring loops, picking the right pre-built tool stack, and running parallel threads for fresh context.

Steal fora checklist for evaluating any coding agent's real day-to-day usefulness beyond raw model benchmarks
20:59concept

The agent loop

  1. Well-defined prompt
  2. Feedback engine (reference examples or a working app)
  3. Success criterion

A three-part structure for getting an agent to self-improve without hand-holding: define what you want, give it something to compare its output against, and state what 'done' looks like — the agent then loops until it clears the bar.

Steal forcode review automation, ad-script generation, thumbnail design, or any task with a comparable reference set
CTA Breakdown

How they asked for the click.

VERBAL ASK
46:05subscribe
“make sure you subscribe to the Riley Brown channel”

Delivered as a quick aside from Ras Mic mid-farewell rather than a dedicated pitch segment — soft, in-conversation ask.

MENTIONED ON CAMERA
00:45channelRas Mic's YouTube channel ↗
FROM THE DESCRIPTION
OTHER LINKSAlso linked in the description.
Storyboard

Visual structure at a glance.

cold open
hookcold open00:00
framework whiteboard
valueframework whiteboard16:06
Sites feature demo
valueSites feature demo37:22
subscribe CTA
ctasubscribe CTA46:05
Frame Gallery

Visual moments.

cold open00:00Frame at 00:38 from OpenAI Merges ChatGPT and Codex00:38Frame at 01:19 from OpenAI Merges ChatGPT and Codex01:19Frame at 01:50 from OpenAI Merges ChatGPT and Codex01:50Frame at 02:24 from OpenAI Merges ChatGPT and Codex02:24Frame at 03:26 from OpenAI Merges ChatGPT and Codex03:26
cold open00:00Frame at 00:38 from OpenAI Merges ChatGPT and Codex00:38Frame at 01:19 from OpenAI Merges ChatGPT and Codex01:19Frame at 01:50 from OpenAI Merges ChatGPT and Codex01:50Frame at 02:24 from OpenAI Merges ChatGPT and Codex02:24Frame at 03:26 from OpenAI Merges ChatGPT and Codex03:26Frame at 03:42 from OpenAI Merges ChatGPT and Codex03:42Frame at 04:29 from OpenAI Merges ChatGPT and Codex04:29Frame at 05:06 from OpenAI Merges ChatGPT and Codex05:06Frame at 05:26 from OpenAI Merges ChatGPT and Codex05:26Frame at 06:04 from OpenAI Merges ChatGPT and Codex06:04Frame at 06:35 from OpenAI Merges ChatGPT and Codex06:35Frame at 07:30 from OpenAI Merges ChatGPT and Codex07:30Frame at 08:03 from OpenAI Merges ChatGPT and Codex08:03Frame at 08:24 from OpenAI Merges ChatGPT and Codex08:24Frame at 08:58 from OpenAI Merges ChatGPT and Codex08:58Frame at 09:39 from OpenAI Merges ChatGPT and Codex09:39Frame at 10:12 from OpenAI Merges ChatGPT and Codex10:12Frame at 11:20 from OpenAI Merges ChatGPT and Codex11:20Frame at 11:24 from OpenAI Merges ChatGPT and Codex11:24Frame at 12:10 from OpenAI Merges ChatGPT and Codex12:10Frame at 12:52 from OpenAI Merges ChatGPT and Codex12:52Frame at 13:12 from OpenAI Merges ChatGPT and Codex13:12Frame at 13:50 from OpenAI Merges ChatGPT and Codex13:50Frame at 14:39 from OpenAI Merges ChatGPT and Codex14:39Frame at 15:08 from OpenAI Merges ChatGPT and Codex15:08framework whiteboard16:06Frame at 16:16 from OpenAI Merges ChatGPT and Codex16:16Frame at 16:49 from OpenAI Merges ChatGPT and Codex16:49Frame at 17:36 from OpenAI Merges ChatGPT and Codex17:36Frame at 18:14 from OpenAI Merges ChatGPT and Codex18:14Frame at 19:04 from OpenAI Merges ChatGPT and Codex19:04Frame at 19:19 from OpenAI Merges ChatGPT and Codex19:19Frame at 19:44 from OpenAI Merges ChatGPT and Codex19:44Frame at 20:38 from OpenAI Merges ChatGPT and Codex20:38Frame at 20:59 from OpenAI Merges ChatGPT and Codex20:59Frame at 21:49 from OpenAI Merges ChatGPT and Codex21:49Frame at 22:25 from OpenAI Merges ChatGPT and Codex22:25Frame at 22:49 from OpenAI Merges ChatGPT and Codex22:49Frame at 23:37 from OpenAI Merges ChatGPT and Codex23:37Frame at 23:59 from OpenAI Merges ChatGPT and Codex23:59Frame at 24:49 from OpenAI Merges ChatGPT and Codex24:49Frame at 25:25 from OpenAI Merges ChatGPT and Codex25:25Frame at 26:09 from OpenAI Merges ChatGPT and Codex26:09Frame at 26:24 from OpenAI Merges ChatGPT and Codex26:24Frame at 27:01 from OpenAI Merges ChatGPT and Codex27:01Frame at 27:44 from OpenAI Merges ChatGPT and Codex27:44Frame at 28:24 from OpenAI Merges ChatGPT and Codex28:24Frame at 29:00 from OpenAI Merges ChatGPT and Codex29:00Frame at 29:30 from OpenAI Merges ChatGPT and Codex29:30Frame at 29:56 from OpenAI Merges ChatGPT and Codex29:56Frame at 30:45 from OpenAI Merges ChatGPT and Codex30:45Frame at 31:32 from OpenAI Merges ChatGPT and Codex31:32Frame at 31:49 from OpenAI Merges ChatGPT and Codex31:49Frame at 32:35 from OpenAI Merges ChatGPT and Codex32:35Frame at 33:05 from OpenAI Merges ChatGPT and Codex33:05Frame at 33:57 from OpenAI Merges ChatGPT and Codex33:57Frame at 34:26 from OpenAI Merges ChatGPT and Codex34:26Frame at 35:03 from OpenAI Merges ChatGPT and Codex35:03Frame at 35:18 from OpenAI Merges ChatGPT and Codex35:18Frame at 36:10 from OpenAI Merges ChatGPT and Codex36:10Frame at 37:04 from OpenAI Merges ChatGPT and Codex37:04Sites feature demo37:22Frame at 37:58 from OpenAI Merges ChatGPT and Codex37:58Frame at 38:49 from OpenAI Merges ChatGPT and Codex38:49Frame at 39:10 from OpenAI Merges ChatGPT and Codex39:10Frame at 39:46 from OpenAI Merges ChatGPT and Codex39:46Frame at 40:21 from OpenAI Merges ChatGPT and Codex40:21Frame at 40:45 from OpenAI Merges ChatGPT and Codex40:45Frame at 41:34 from OpenAI Merges ChatGPT and Codex41:34Frame at 42:10 from OpenAI Merges ChatGPT and Codex42:10Frame at 42:51 from OpenAI Merges ChatGPT and Codex42:51Frame at 43:21 from OpenAI Merges ChatGPT and Codex43:21Frame at 43:57 from OpenAI Merges ChatGPT and Codex43:57Frame at 44:33 from OpenAI Merges ChatGPT and Codex44:33Frame at 44:55 from OpenAI Merges ChatGPT and Codex44:55Frame at 45:34 from OpenAI Merges ChatGPT and Codex45:34subscribe CTA46:05Frame at 46:39 from OpenAI Merges ChatGPT and Codex46:39Frame at 47:24 from OpenAI Merges ChatGPT and Codex47:24
Watch next

More from this channel + related breakdowns.