Modern Creator
Riley Brown · YouTube

I Spent $10,000 Testing AI Agents (Here's the Stack I Actually Use)

A creator with 2 million followers maps his whole agent stack onto one grid: Claude for making things, Codex for running the business, Composio as the shared toolbelt, and GrokBot as the Twitter specialist.

Posted
yesterday
Duration
Format
Tutorial
educational
Views
2.9K
106 likes
Big Idea

The argument in one line.

Pick agent platforms by job rather than loyalty: talk to one orchestrator per platform, let it dispatch local or cloud sessions, and keep plugins and skills in a shared layer so switching platforms costs nothing.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…
  • You already pay for both Claude and ChatGPT and keep asking yourself which one to open for a given task.
  • You run a solo or small content business and want agents to handle email, Slack, uploads and scheduling while you are away from the desk.
  • You want to control coding and editing sessions on your own computer from your phone.
  • You have set up the same integrations (Gmail, Notion, Slack, Vercel) in three different agent apps and are tired of repeating it.
SKIP IF…
  • You want a step-by-step setup tutorial; this is a tour of how one person splits the work, with few configuration details.
  • You use a single chat assistant occasionally and have no need for parallel sessions or computer control.
  • You need a solution for shared agent memory; the video names it as an open problem and does not solve it.
TL;DR

The full version, fast.

The stack is organized by job, not by brand loyalty. Claude Code, driven through Claude Projects, handles anything that produces an asset: apps, front-end changes, slide decks that explain dense documents like Anthropic's S-1, design boards, and full video edits. Codex, driven by voice through ChatGPT's Dot agent, handles computer use and the administrative work: uploading YouTube drafts, reading and answering email on a walk, Slack and calendar. In both cases the user talks to a single orchestrator that decides whether to spin up a local session on the laptop or a cloud session. Composio sits underneath as a shared plugin and skill layer so any new agent platform inherits every integration at once, and GrokBot is called by webhook from the other agents only for Twitter access. Shared memory across platforms is the missing piece.

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 – 00:50

01 · Cold open: the $10,000 test

Too many agent tools; the promise is a settled stack of three platforms plus one tool in the middle that shares plugins and skills.

00:50 – 05:34

02 · Claude Projects as the coordinator

A coordinator session per project dispatches local or cloud Claude Code sessions. Clipping runs on the laptop while a best-practices artifact is built in the cloud, all controllable from the iPhone app and by voice.

05:34 – 08:45

03 · Building apps

An internal tool called Native Note replaces about ten SaaS apps. A phone prompt restyles the Kanban view, Opus 5.5 makes the changes locally and deploys. Claude wins on front-end work.

08:45 – 11:30

04 · Design and slides

Anthropic's S-1 becomes a visual slide deck in three prompts from the phone. Claude's design mode produced the original Native Note design board, which can be pushed into the app.

11:30 – 14:37

05 · Video editing and V2

Two recent videos and an animated explainer of a Karpathy tweet were edited by AI. The creator's free, open-source V2 app adds version control for video edits; waitlist pitch.

14:37 – 15:49

06 · Sponsor: Superhuman

Superhuman Mail and Docs connectors in ChatGPT turn video-deadline emails into a content calendar.

15:49 – 18:46

07 · Dots with voice

ChatGPT's Dot is a Codex orchestrator. On a pre-gym walk, a voice call spins up a local Codex session to update Native Note, a cloud research report, and a Notion status check.

18:46 – 20:09

08 · Computer use

When a video is final, Dot has Codex upload it in Chrome, write the description, pull sponsor links from email and save a draft for review.

20:09 – 21:15

09 · How the work splits

Claude for assets and learning, Codex for personal-assistant and admin work. Claude Duet is passed over because it cannot run local sessions.

21:15 – 22:25

10 · Email and Slack on a walk

The agent auto-archives noise, reads the remaining 20 to 25 emails aloud by importance, and drafts replies by voice.

22:25 – 27:01

11 · Composio: shared plugins and skills

About 30 connected apps, a custom Native Note tool and 20 skills live in Composio, so any agent platform inherits them. Instant bundles paid APIs like Exa, ElevenLabs and FAL behind one key.

27:01 – 28:00

12 · Memory, the unsolved piece

Every platform keeps its own memory and none of it transfers. Supermemory through Composio was tried and didn't update often enough.

28:00 – 29:50

13 · GrokBot via webhook

GrokBot is never opened directly. A webhook skill lets Claude Code or Codex hand it Twitter research and DMs.

29:50 – 31:28

14 · Recap and sign-off

The finished grid: Claude for assets, Codex for business, Composio in the middle, GrokBot for Twitter. Expect it to change in two months.

Atomic Insights

Lines worth screenshotting.

  • The most useful agent setup is one orchestrator you talk to, which decides on its own whether a task needs your laptop or can run in a cloud sandbox.
  • An orchestrator that cannot spin up sessions on your own computer is a dealbreaker, because so much real work lives in local files.
  • Claude Opus 5.5 leads on front-end design and visualizing data, while GPT 6.1 Sol is frontier-level but weaker at front-end polish.
  • Turning a dense document like an S-1 filing into a slide deck is a faster way to learn it than reading it, and three prompts were enough.
  • Voice is the highest-minutes use of AI here: an hour of walking while talking to an agent replaces desk time for email and Slack.
  • Around 90% of email can be handled on a walk when the agent pre-filters junk and reads the remaining 20 to 25 messages aloud in order of importance.
  • Let an agent with computer use prepare the YouTube upload as a draft, including sponsor links pulled from email, and keep the publish click for yourself.
  • Split the work: Claude for making assets and learning, Codex for running the business through computer use, email, Slack and texts.
  • A shared plugin layer like Composio means testing a new agent platform takes one connection instead of re-authorizing 30 apps.
  • Neither Claude nor ChatGPT ships a native YouTube plugin that can change your channel, but a third-party integration layer can supply one.
  • Keeping skills in one central place stops you from copying the same skill files between Claude Code, Codex and every other agent.
  • A webhook skill lets your main agents delegate to a specialist agent, so you get GrokBot's Twitter access without ever opening its interface.
  • Memory is the unsolved layer of a multi-agent stack: each platform remembers you separately and none of it carries over when you switch.
  • Any agent stack you settle on today will likely look different in two months, so optimize for how cheaply you can switch.
Takeaway

Assign each agent the job it does best.

WHAT TO LEARN

One orchestrator per platform, a clear make-versus-run split between Claude and Codex, and a shared integration layer turn four agent platforms into one stack you can swap pieces of.

02Claude Projects as the coordinator
  • Talk to a single coordinator per platform and let it decide whether each task runs on your laptop or in a cloud session, instead of choosing yourself.
  • Keep your computer on and you can launch and monitor local coding or editing sessions from your phone while away from the desk.
03Building apps
  • Give each app you maintain its own project so a quick phone prompt carries that app's context and can ship the change end to end.
  • Use Claude for front-end-heavy builds, since it currently produces cleaner interface design than the competing frontier model.
04Design and slides
  • To understand a dense document like an S-1, have an agent turn it into a visual slide deck and iterate on it in a few prompts.
  • Design boards with several options let you pick a direction visually and then tell the agent to implement it in the real app.
05Video editing and V2
  • AI can carry a full video edit when you iterate in versions, and tracking the prompts behind each version makes the result repeatable.
07Dots with voice
  • A voice agent on a walk can start coding sessions, research reports and status checks in parallel, turning dead time into output.
08Computer use
  • Let computer use prepare uploads as drafts, with descriptions and sponsor links gathered from email, and keep the final publish for yourself.
09How the work splits
  • Route asset creation and learning to Claude, and route administrative actions like calendar, email, Slack and texts to Codex.
  • Reject any orchestrator that cannot reach your local machine if your work depends on local files and apps.
10Email and Slack on a walk
  • Have the agent pre-filter and auto-archive email so only the 20 to 25 messages that matter get read aloud and answered by voice.
11Composio: shared plugins and skills
  • Connect integrations once in a shared layer like Composio so every new agent platform gets all of them through a single plugin.
  • Store skills centrally rather than inside one agent, so moving from one platform to another doesn't mean rebuilding them.
12Memory, the unsolved piece
  • Plan for memory to be the gap: each platform remembers you separately, so budget for repeating context when you switch.
13GrokBot via webhook
  • Reach specialist agents through a webhook skill so your main agents can delegate niche work, like Twitter research, without you changing apps.
Glossary

Terms worth knowing.

Claude Code
Anthropic's agentic coding tool that can read and change files and run commands, either on your own computer or in a cloud sandbox.
Codex
OpenAI's agent for coding and computer tasks, which can run sessions on your machine or in the cloud.
Orchestrator agent
A main agent you talk to that does little work itself and instead starts and tracks other agent sessions, passing each one its task.
Local vs cloud session
A local session runs on your own computer with access to your files and apps. A cloud session runs on a temporary remote machine, so your computer can be off.
Computer use
An agent's ability to operate a computer like a person does: open a browser, click, type and fill in forms, usually guided by screenshots.
Composio
A service that hosts connections to many apps (Gmail, Slack, Notion, GitHub and more) plus reusable skills, which you plug into any agent platform at once.
Skill
A saved set of instructions, and sometimes scripts, that teaches an agent how to do a specific recurring task the same way every time.
Webhook
A URL that triggers an action when another program sends it a request. Here it lets one agent hand a task to a different agent.
S-1
The registration statement a US company files with the SEC before an IPO, disclosing its finances, risks and business details.
Net revenue retention
How much revenue a group of existing customers generates now compared with a year earlier, including upgrades and churn. Above 100% means existing customers spend more over time.
Resources

Things they pointed at.

01:04toolClaude Code / Claude Projects
05:34toolCodex
05:34toolDot (ChatGPT personal agent)
06:24productNative Note (creator's internal tool)
08:45linkAnthropic S-1 filing
14:37productSuperhuman Mail and Docs (sponsor)
17:34toolGPT Pages
20:54toolClaude Duet
22:40toolComposio ↗
23:20toolExa
25:30toolInstant (single-API key bundle)
25:30toolFAL
27:22toolSupermemory
28:00toolGrokBot
Quotables

Lines you could clip.

18:46
“It is like talking to a human that can use codex for you.”
one-line definition of a voice orchestrator→ TikTok hook↗ Tweet quote
20:54
“Claude Duet cannot spin off local sessions running on your computer. And that is a non-negotiable for me.”
clear opinion with a reason→ IG reel cold open↗ Tweet quote
21:15
“Like 90% of my emails is I pull out my phone in the morning, put in my AirPods and go for a walk.”
specific, relatable habit with a number→ TikTok hook↗ Tweet quote
05:13
“I created this presentation this morning using Claude projects from my phone, walking back from the gym.”
proof of the phone-first workflow→ newsletter pull-quote↗ Tweet quote
27:01
“The same way I've centralized my skills in plugins, I want to centralize my memory.”
names the open problem in the space→ 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.

metaphoranalogystory
There are so many AI agent tools out there and it's starting to get incredibly overwhelming. So I spent $10 ,000 testing all of the different AI agent tools in pursuit of finding the greatest agent stack that exists today. And I've settled on what I think is the best stack of tools that I use for administrative tasks, running my business and growing my following to over 2 million followers across all platforms.
So today we're talking about my full AI agent. stack and all of the main use cases for each one. I think you're going to be surprised at how I use these three agent platforms.
It's a little bit different than how most people use them. And we're also going to be talking about a tool that sits at the center of my agent platforms that makes it easier to switch between them. And it allows me to share plugins and skills.
Okay, so many people ask me, when do I use Cloud Code and when do I use Codex? And so for these first eight use cases, that's exactly what we're going to be talking about. And so the first way that I use Cloud Code.
that I think everyone should 100 % be using is Claude projects. And so I made a video about a week ago where I said that you should not be using normal Claude and that you should be using Claude code because it's so much more powerful and it has access to all of the files on your computer and running Claude code locally on your computer allows you to use prompts like this.
Hey, there's a folder in my downloads called October 26th vlog. Please edit the video and post it on YouTube. And Cloud Code can end to end complete this full task.
But there's also another setting in Cloud Code that you should 100 % be aware of, right? Not only can you run it on your computer, you can also run it. in a cloud session.
It'll actually spin up a little computer, do anything that Cloud Code might want to do. It's just not running locally on your computer. And for a lot of tasks, that's really useful because you don't actually need your computer for all of the different tasks.
And so if you think this is confusing, that's perfectly OK, because cloud projects will actually manage this entire process. On the left here, you can see all of the cloud projects that I've created. And here I've opened up editor.
And so this main session is just a coordinator session and it will spin up different Claude sessions, whether it is a local session or a session on my computer. So I can say something like this. Hello, I need you to please look at my latest videos in my downloads and I want you to make three clips and edit these locally on my computer and upload them to drive and give me a link to all three.
So I can fire that prompt and look at this. It was actually sent to an existing thread. So sent to a thread.
Here's a vertical clips from side channel video. And if you see right here, we can see that this is running locally on my computer. And so if I click on threads, I can actually see all of the different threads that are running right now.
Some of which are on the computer and some of which are in the cloud. Like this one is in the cloud and all of these threads are running in the cloud. And so right now we see that my local session is working.
And if I click on this, we can see that the prompt that I entered here was automatically passed to this local session. And now cloud code is working and it's working locally on my computer, but I could say, Hey, while that's working, I actually want you to look up best practices and create a little artifact for this. Like I want to create like a little webpage artifact that I can use and send to my team on best practices for clipping long form videos into short form.
This can just be done in the cloud. And so I can fire this prompt in the main thread or the coordinator thread called editor, which is a cloud project. And here, look at this.
It says right clip. So it just created a new session. Notice here, there's no little computer icon.
This one's actually running in the cloud. And this is really useful because I can do all of this from my phone. And so if we open up the Claude iOS app and I click on this side panel right here, and then I can see all of my projects right here.
So if I click on editor, I can see the sessions that are running right now. And so as long as my computer is on, I can fully control my computer using Claude projects right here. We can see that the web one is done.
So let me click on this. It should have sent me. a page.
Here is the artifact that it sent me and I can very easily see it. This is a shareable artifact that I can share with my team. And this is pretty cool.
It's pretty easy. It shows the framing and safe zones, which is pretty cool. And one of my favorite things that I do with cloud projects is I will come down here to the bottom.
I will press this button, open up the artifact. And then I'll just speak out the changes that I want to make. I actually want this to be better format and just make the design a lot better.
And the top table, change it, please. And it allows me to just record my voice and I can create artifacts. And that's exactly how I created this presentation right here.
I created this presentation this morning using Claude projects from my phone, walking back from the gym. I had this idea for this video. I said, create this slide deck and I have a very specific skill that I use for it.
And when I got back, it was done. And so I use cloud projects to coordinate my cloud sessions, whether that's cloud sessions or my local sessions that are running on my computer. And this is very similar to how I use Codex or their new dots feature.
You can very similarly, you can speak to your dot and you can get it to coordinate codex sessions. And we'll talk about that a little bit later. But now for the next three items on this list, I want to talk about how I use cloud code or why I use cloud code and what sets cloud code apart from codex.
And so the first one is building apps. Some of you will recognize this. I talk about this in a previous video, but I have my own internal tool that is a mixture of like 10 different SaaS apps that I use for everything.
And that includes Excalibur. Right. We have all of my Excalibur documents.
This is my own clone of Excalibur. All of it is stored in my own internal tool. I also have my mind built in, which stores all of my bookmarks.
I have all of my YouTube analytics and I even have my own video editing platform that. I will talk about a little bit later. And in order to manage this entire site, I have this thing called native note.
So this is called native note. And this is where I build this app and make changes to this app. And so on my phone, I could very easily go to projects.
And we can move to my agent, my native note project and say, hey, what I need you to do right now is I actually don't like how on native note it has my pinned videos. Can you please remove that? And then also move the Kanban title to the top.
We also don't need that little subtitle or this like horizontal line. And please make the filter by title thing pinned to the left instead of on the right. And yeah, please make the background.
have a little bit more texture to it, almost like it's like a whiteboard. And I can fire that prompt off. Notice how I didn't specify whether I want it to be local or in the cloud.
It will decide whether it needs my local computer or not. And we've been working a lot. And see here, I've sent the Kanban cleanup to this thread, which I can click on.
And here it is now building. And here you can see it received the message from the coordinator and it'll begin working. And so you can see here that Opus 5 .5 is running on a cloud code session that is running locally on the computer.
And it is done with three out of the four tasks. It has made all those changes that we just asked. And now it's just waiting to deploy.
So it actually shows up in native note. OK, so now it is done and it should be live. And so if we go to native note, we press command R and we refresh it.
We see that we now have a. new app here it is updated and so anytime i want to change to the app whether i'm on my phone or my computer i can just ask claude projects all i need to do is go to the native note project just ask it And it'll figure out how to do it.
And so Claude Opus 5 .5 right now, in my opinion, is at least for the things that I'm doing, it is the best at coding. And so it's best at creating apps, especially if it is very front end focused. It's just very good at front end.
And GPT 6 .1 Sol is also frontier and it's amazing. It's just not the best at like front end design. And so that is why I use Claude code specifically projects to build apps.
And so the next thing that I use Claude for, specifically Claude Projects, is I use it for design and slides. Yesterday, Anthropic released their S1, which is a required document outlining a ton of different things that's required before an IPO. And on my phone, using Claude Projects, specifically I was using my RIC.
Claude project, which is basically like Claude Duet, which is my own little personal agent. But I just think projects is better. I talked to this agent on my phone.
I was like, yo, this S1 came out. It's very text heavy and data heavy. I need to properly visualize it, organize it how we normally organize my videos and create a slide deck.
And like, this is what it created. It shows all of the data in this very easy to understand format where the money goes. I could look at their revenue, their operating costs.
And what I've noticed. is Anthropix models, especially Opus 5 .5 is just incredible at visualizing information. Take any information in the world, a PDF that's super data heavy or an Excel sheet and just tell it to turn it into like this visual thing.
Like here it says, here is the amount of customers who spend around 250. Or 250 spend 10 mil or more. 1900 people spend a million or more.
And it just creates these little visual diagrams here. It shows revenue attention. So a customer spent a dollar a year earlier.
That same customer spends nine dollars now, which is just absolutely unheard of. And it goes through all of its compute. It takes me through everything, all of their risks.
And then also it just has really good organization. Many of you who watch my videos know that I have this working outline that gets filled in as we go through the video. And so it's super cool.
And then I had it do a data section at the end. And so we can look at all the raw numbers and it created this in three prompts. I just went back and forth on my phone saying like, Hey, can you change this?
Can you change this? And this is how I learned about everything on the S1 was just with this slide deck. And again, this is not just slide decks too.
This is also their cloud design mode, which is integrated with cloud projects. This is actually where I originally designed native note. As you can see here, this was how I kind of originally designed the entire app.
And I just iterated with Claude. I think I did this one on my computer, but you can create these designs on your phone. Anything you do on your computer, you can now do on your phone with Claude projects, which is why I use it so much.
Just say, Hey, create a Figma style design board, create, you can create as many options as you want. And then you can just select the option that you want. You can say, Hey, push this to prod or implement these designs into the app that we're creating.
And it can. spin up a local session to do that. It may design it in the cloud and may implement it using your computer and you don't even need to like know that you can just let your agent do its thing.
And so I use it for slide decks and then creating Claude design. Now here's for the most fun one and the most viable use case right now. And that is video editing.
And so I've been using Claude nonstop to edit my videos. In fact, these two videos right here. this how smart people are using AI and how I've been using Claude duet, both of these videos were fully edited by AI.
Now there was some iteration, but they were edited by AI. And so using Claude projects was also how I've been creating these animated explainer videos with a very specific style. And so as you can see here, I created with Claude and we can actually see which model So yeah, here are my prompts to create version one of this video.
I used these prompts right here, and it created this video, which explained one of Karpathy's tweets. With LLMs at all. Around 1 to 2 billion, about 20%, are casual and infrequent users of free -tier ChatGPT -like products.
This group sees derpy chatbots and treats them a bit like a better Google search, a writing aid, etc. Maybe an agent tries to book you a flight. I have non -tech friends who, reasonably, in my opinion, say they have not much use for it at all.
Even many professionals outside of math and code are in this tier. And then that was just version one. This was actually version two of the video.
I basically made some changes with these prompts using Cloud Opus 5 .5 on projects and it made the video much better and it was more zoomed out. And I posted this video on Twitter and it actually got 2000 likes and 160K views. And many people were commenting, asking me about the skills like, yo, share how you made it.
And do you mind sharing the prompt that led to the graphics or is it just Opus 5 .5? choice. And by the way, this little app that you're seeing right here, this is V2.
This is an app that I'm working on that will be free and open source. It is the GitHub for video editing and it will allow your agents to create versions of the app that you create and you'll be able to track all of the prompts. and all of the different files that you use to create each video.
And people on your team will be able to fork your video. You'll be able to share your videos and make them public. And yeah, I'm really excited about this project.
And it is video editor agnostic. You can use any stack. You can use any agent, anything.
All you do is copy a prompt, give it to Codex, and it will create version control for your video. And all you need to do is go to v2 .video and you can sign up. for the wait list.
As soon as it's ready, we will send access to everyone who puts their email right here. I'm really excited about this. This is free and this will always be free.
This is the GitHub for video editing. Working on it right now with my team. Okay, so in summary, with Cloud Code, it's just incredible at creating things.
And I think the best orchestrator of creating things is Cloud Projects because you can create different projects, one for each of the apps that you're building. You can create an editor agent and a designer agent, and it's really useful. But there are also things that Codex is better at than Cloud Code.
So now we need to move down the stack and we're going to get to Codex. What if your inbox could become a plan you actually use? This part of the video is sponsored by Superhuman.
As you guys know, I use AI tools for basically everything, and I already use Superhuman for email. but pulling together deadlines and feedback still takes work. Superhuman Mail and Superhuman Docs are part of the Superhuman Productivity Suite.
Mail is the email app. Superdocs is a collaborative workspace for documents, tables, and project trackers. Both have connectors for the AI tools you already use.
Here's a workflow I wanna show you in ChatGPT. First, I connect the Superhuman Mail and ask, find the emails about my upcoming videos, What are the deliverable deadlines and latest feedback?
Then using SuperDocs connector, it will turn that into a content calendar with a next action for every video. Flag anything that's missing. Now I have a plan I can review against the actual emails.
When something changes, I can ask the AI to update the document. Try SuperHuman connectors today. You can sign up for SuperHuman using my link in the description.
all day, every day basically is I use dots with voice. So if we were to measure how I use AI in minutes, this is the number one way I use AI. And I've gotten into a routine.
I go to the gym in the morning. But I leave an hour and a half early and I walk around for an hour before I go to the gym talking to my dot. So dot is the personal agent platform that is in chat GPT.
And I think the correct way to think about dot is it is a codex orchestrator. The same way that cloud projects can delegate to cloud code sessions, dot can delegate to codex sessions. And the same way that cloud code can spin off cloud sessions and sessions running on your.
computer dot can do the same thing with codex so i can call dot hey riley good to hear you what's on your mind hey pal i need you to check the recent changes that i've made to native note can you spin up a new codex session investigate it and i actually want you to add that same background that we added to native note to the other pages like to the background of the list view and in the calendar view as well can you spin up a new codex session and then push your changes to prop On it.
I'll open a fresh Codex session and dig through those recent changes now. I'm getting a new session ready for that background update now, including a prod push once it's checked. Thank you, buddy.
Of course. And look at this. It shows Unify Native Note backgrounds.
If I click on this right here, we can see that it's spun up. Thanks for waiting. The new Codex session is running on your MacBook Pro and inspecting Native Note.
The changes and production push are still ahead. I'll keep an eye on it and let you know as soon as there's something to review.
OK, that's great. Yep. I can see the codec session running in the sidebar while that's running.
Can you actually spin up a new cloud session real quick and just do some research for me on how people are video editing with AI? I want like an in -depth report and you can create a pages, you know, in GPT space, you have the new page feature. I want you to create pages.
I want diagrams of exactly how this is working.
And yeah, that's what I kind of want to know is like, what are the strategies people are using? What are the types of skills people are using? Put it in the GPT pages.
And yeah, spin this out and put codecs using the GPT -6 Sol model to do this. Yep, I got it. I'll start that cloud session with GPT -6 Sol for the in -depth report on AI video editing with pages and diagrams showing how the tools and skills fit together.
And meanwhile, I'll keep an eye on that native note session so we're not doing anything twice. Amazing. Yeah, cool.
And then actually, once you're done with spinning up that session, can you do you mind just give me a little report here in the chat on like what videos I'm working on in Notion, my long form videos? Sure, I'll pull that up. I'll check Notion for your current long form videos and put a short status report here in chat.
Thanks, pal. So that's why I use dots with voice. It is like talking to a human that can use codex for you.
And so codex is like very, very good at a lot of things. And the next reason I use dots and codex as like one of the main tools that I use is it has incredible computer use. In fact, whenever I edit a video and I really like the video and I get to my final version, I just ask my dot.
I said, hey, the video is ready. Post it on YouTube. And what it'll do is it'll analyze the transcript of the video.
It will go through in Chrome, upload the YouTube video there, fill out all of the details, get the description ready. Look at my emails. You know, if it's a sponsored video and I need to include like a link in the description, it'll add it there automatically.
And then it will save it as a draft. And then my dot will just message me like your dot, your draft is created here. Go ahead and look at it.
And when you're ready, you can post it. It's incredibly useful and computer use. with Codex is amazing.
Next time you're on your phone and you leave your laptop open, just go to your dot and be like, hey, spin up a new Codex session and ask it to do anything on your computer and have it take screenshots as it's doing it. And you can like kind of see how your agent uses your computer. And it's basically like having an agent that can fully control your computer.
And so that is one of the main reasons. I also use Codex a lot for business tasks. That's how I kind of split.
my usage between cloud code and codex cloud code is creating assets for my content business slash when i want to learn anything like in the case of the the anthropic s1 i was just wanted to learn that's why i had it create slides and i do that for basically everything that i want to learn so if i'm creating an asset i will basically have cloud code do it because it is just really good at visual design, and also video editing.
And Codex I use for my personal assistant administrator, and it keeps track of my calendar, my emails, all the messages in Slack. That's how I use Codex. And I generally just speak with Dot because Dot can spin up Codex sessions.
So in both of these cases, with Dots and projects, I'm speaking directly to an orchestrator. And so Anthropic, what I think they're trying to do is they're actually releasing a new product called Claude Duet. Most people don't have access, but it is their version of Dots and it is a standalone app.
I prefer projects now because Claude Duet cannot spin off local sessions running on your computer. And that is a non -negotiable for me. I must be able to do it for me to use that agent.
frequently. And so that is why I'm using these two apps as my orchestrator. And the final use case that I've been using for Codex is email and Slack.
I mentioned it before. I can't really show examples because all of the data is sensitive. The way I do all of my emails now, and I'm not kidding, like 90 % of my emails is I pull out my phone in the morning, put in my AirPods and go for a walk.
And I just talk to the agent and be like, okay, you know, Every morning there's probably, my AI is already set to filter out all of the emails that I don't need to see. It'll just auto archive them.
And then there's probably like 20, 25 emails that I need to hear. And I just talked to my dot. I said, in order of importance, please read me my emails.
And then my dot will talk to me and be like, this one's from this person. And they said this, and then it'll read the email. And I'll be like, okay, dot, I want you to respond with something like blah, blah, blah, blah, blah.
And then it'll read it out to me and I'll be send it. And then I'll do that on a walk. I love walking.
I hate sitting in front of a computer and doing emails. I do all of my email and responding to Slack in little breaks in the day when I go for a walk. When I can go for a walk, I do.
In fact, I can't wait to do it after this video and talk to the dot and I can actually get things done while walking around. It's amazing. Okay, so now what I want to talk about is a little bit different of a tool that sits at the center of all of my agent platforms.
And this tool is what allows me for my agents to share skills and plugins. And by the way, this is not sponsored at all. It's just an incredibly useful tool.
And that is Composio. And so Composio allows me. to centralize my skills and plugins.
Many of you guys have seen this screen before in the ChatGBT app, and there's an equivalent one in the Cloud Code slash Cloud app where you can install these plugins. And they're really useful because plugins allows your agent to actually use all of these tools. In ClickUp, I store all my brand deals with my management company.
In Convex, that's where I store all the data for all my apps. In Gmail, that's where I have all my email. I have my calendar.
I have my superhuman email, which I actually use. more than the Gmail one. I have my Notion where I store a lot of useful things.
Vercel where I host the apps. Exa where I do, it has a better web search than the ones that are built into these agent platforms. And then a lot more.
And it would be really annoying that if every time I wanted to test a new agent platform, I would have to re -sign in to all of these. Well, one plugin that you can add to all of your agents is the Composio plugin. And this is Composio.
Once you create your account. All you have to do is go to apps and you have literally every plugin you've ever created. Claude and GPT have no built -in YouTube plugin where you can plug it into your actual YouTube account where it can change things on your YouTube.
They have that on Composio. And so all you have to do is plug in your Composio and then you get access to all 30 of these apps that I've added. And you can create your own custom tools.
like I did here with my native note tool, right? I have an internal tool and I created a native note app on Composio. And all I did is ask my agent to create it.
Now, all of the agents that I connect to will be able to use native note. It will have all of the instructions and it's really easy to use. And so remember with dot, we were asking it to create both of these sessions right here.
If we navigate to Composio and we go to home and we scroll down, we can see that it was actually using. The Composio skills the whole time, the way that it uses Slack is through Composio. The way that it uses Gmail, it's not even using the ones baked into Codex.
It's using the ones on Composio. And so I'm able to just give my agent the ability to use all of the plugins. And so that means let's say tomorrow Microsoft came out with the best agent platform and it was powered by their new model, Microsoft Satya 4 .0.
And this was insane. And they had all these different. plugins that you could add.
All I would need to do is go to this new agent platform and plug in my Composio and it would have access to all of my plugins. Additionally, they have instant. And I'm telling you, this feels sponsored.
I promise you this is not sponsored. This is just genuinely useful. This is Instant.
And so what Instant does is it basically takes all of these useful API keys that you might want access to, like Tavoli, which is search, Exa, which is search. It also has like 11 labs, FAL, which has every popular video model for cheaper. It's much cheaper than Higgs field to use.
You can just plug it into FAL and get all of the creative video. models and image models. What it does is it basically sets this up to a single API.
So it'll charge you for the amount that these APIs cost, but you don't actually have to go in and set up your API key. It just automatically connects to all of them. And so the way that I think of Composio, right, it's kind of an agent.
It allows me to very easily switch from agent to agent platform because I've centralized all my skills and plugins. And by the way, you can also create skills. I think I only showed you the apps that I had built.
So in Composio. You have your apps and you also have skills. And then I have mine at my email.
And here are 20 skills that I use quite frequently. I don't want to have to set up skills on Claude Code and then tell Claude Code to send the skills to Codex or send them to Grokbot. I just centralize them.
All of my skills are in one place. And all I need to do is connect the Composio integration and it has all of these skills. And the next one is actually one that.
It's something I'm working on right now, and I actually don't know how to solve the problem. And for anyone who solved this problem, I would love for you to let us know in the comments. And that is memory.
The same way I've centralized my skills in plugins, I want to centralize my memory. Codex, Cloud Code, Grokbot, Muse, Cloud Duet, all of these platforms keep a memory on you. But when you switch from platform to platform, the memory does not persist.
And so I'm trying to figure out a workflow that is relatively simple that allows these agents to centralize my memory. I tried connecting super memory to Composio and that was okay. I just couldn't figure out how to get it to update its memory more frequently.
So I could very seamlessly switch from Claude code to Codex and it could just pick up with all of that context in mind. It would actually have memory of all the things that I'm. doing.
So that one's more of an open question because this is the thing that I need to like complete my stack. I need to figure out memory. So let me know in the comments.
Okay. Finally, we need to talk about grok bot. Now I don't actually go to the grok bot UI ever.
Really. I mostly just use dots and cloud projects. That's how I interact with agents, but I want to talk about two things.
First of all, Grokbot has Twitter access. And if you're a content creator, it's incredibly useful to get an AI to scour Twitter to find links. Maybe you're looking for B -roll videos to put in your news content.
It's really good for doing research and being up to date, especially on tech news. This is where the smart people are talking about tech, period, on Twitter. And because you can connect Grokbot to Twitter, and then they also have a feature called Webhook, you can allow Cloud Code and Codex to send a request.
to grokbot. Let me show you what I mean. So if we go to Claude, and this is a local Claude code session, I can type in grokbot webhook.
Hey, I need you to research Claude design, not Claude design, rather Claude duet, their new platform. Can you please get grokbot to do some research on this and create a little report based on this. And so what this is going to do is this is going to use a skill that I've configured called grokbot webhook, which is connected to if we open up grok.
Grok bot we can see here that it has a Claude code webhook skill So when a webhook fires it will begin working and you can see here that it's automatically Starting it is chipping away. It is doing research It sent a request to Grokbot.
So that means I don't need to use three different agent platforms. I can basically just interface with my favorite agent platforms, which is Codex or Dots and Claude with Projects, and they can send requests to Grokbot. I don't actually use the Grokbot UI.
I just basically have my agents delegate to Grokbot because Grokbot has full access to Twitter. It's really useful. And this also connects to your account.
So I could have asked Cloud Code to check my Twitter DMs and it would have just delegated that to Grokbot. Very useful. And there you have it.
I am a content creator and I'm starting a company called Agent Native. And these are all of the tools that I use. I use Cloud Code, specifically projects, to coordinate all of the assets that I need to create.
Apps, designs, slides, video editing. I use Codex for its computer use. All business and coordination tasks, which are just such a high percentage of my business, is just coordinating with other people through email, Slack, and otherwise.
And Codex can access my text messages as well, which is really useful. That means I can access my messages through voice with dots. And that's kind of the main breakdown.
And then obviously plug this into all of the other tools that I use. And the way that I centralize that is through Composio. And then Twitter is such a large part of my business, whether that's doing research or posting on Twitter.
It's one of the main accounts. I connected to Grokbot because it has full access to my Twitter account and doing research on Twitter. So this is like 90 plus percent of the things that I do.
with my agent stack and as this evolves i'll probably remake this video once every two months because it'll probably be different in two months because the agent platforms will evolve meta might release a new model google might release a really good agent platform it's really really cool and who knows someone might release a better composio competitor and so anyway thank you guys so much for watching this video i hope you guys got a lot of value if there's something that you want me to create or a video you want to see soon, let me know in the comments.
I'll see you here for the next one.
The Hook

The bait, then the rug-pull.

After $10,000 of testing, the answer to "Claude or Codex?" turns out to be both, each with its own job. The video fills in a four-row grid one cell at a time, and the most useful part isn't any single tool. It's the pattern of talking to one orchestrator and letting it decide where the work runs.

Frameworks

Named ideas worth stealing.

29:50model

The full agent stack grid

  1. Claude Code: Projects, build apps, designs and slides, video editing
  2. Codex: Dots with voice, computer use, business tasks, email and Slack
  3. Composio: centralized skills and plugins, my skills, memory (open)
  4. GrokBot: Twitter access, webhook skill for any agent

Each platform gets a row of jobs it is best at, with a shared integration layer in the middle so the agents can trade work and tools.

Steal forauditing your own AI tool spend by job instead of by vendor
02:13concept

Orchestrator, then sessions

Talk only to a coordinator (a Claude Project or ChatGPT's Dot). It routes each request to an existing thread or spins up a new local or cloud session, so you never pick the environment yourself.

Steal forrunning parallel agent work from a phone
20:09concept

Make vs run split

  1. Claude: create assets and learn things
  2. Codex: administer the business through computer use, email, Slack, calendar and texts

A simple rule for which assistant gets a task, based on whether the output is an asset or an action.

Steal fordeciding which subscription to keep for which work
CTA Breakdown

How they asked for the click.

VERBAL ASK
13:15product
“All you need to do is go to v2.video and you can sign up for the wait list.”

Placed mid-video, right after the video-editing demo had already shown V2 on screen, so the pitch lands on something viewers just watched.

MENTIONED ON CAMERA
Storyboard

Visual structure at a glance.

open
hookopen00:00
empty stack grid
promiseempty stack grid00:35
Projects cell
valueProjects cell04:43
Native Note Kanban
valueNative Note Kanban07:52
S-1 slide deck
valueS-1 slide deck09:26
V2 video versions
ctaV2 video versions12:35
Dot voice call
valueDot voice call17:42
the missing center
valuethe missing center22:25
Composio dashboard
valueComposio dashboard23:36
GrokBot webhook
valueGrokBot webhook28:43
finished grid
ctafinished grid30:17
Frame Gallery

Visual moments.

One-click upgrade to your Google

Get more breakdowns in your search results

Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.

Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
Watch next

More from this channel + related breakdowns.