I Spent $10,000 Testing AI Agents (Here’s the Stack I Actually Use) — Transcript
Full transcript
- 0:00There are so many AI agent tools out
- 0:02there and it's starting to get
- 0:03incredibly overwhelming. So, I spent
- 0:06$10,000 testing all of the different AI
- 0:09agent tools in pursuit of finding the
- 0:12greatest agent stack that exists today.
- 0:15And I've settled on what I think is the
- 0:16best stack of tools that I use for
- 0:19administrative tasks, running my
- 0:21business, and growing my following to
- 0:23over 2 million followers across all
- 0:26platforms. So today we're talking about
- 0:28my full AI agent stack and all of the
- 0:30main use cases for each one. I think
- 0:33you're going to be surprised at how I
- 0:34use these three agent platforms. It's a
- 0:37little bit different than how most
- 0:39people use them. And we're also going to
- 0:41be talking about a tool that sits at the
- 0:44center of my agent platforms that makes
- 0:46it easier to switch between them and it
- 0:50allows me to share plugins and skills.
- 0:52Okay, so many people ask me when do I
- 0:55use cloud code and when do I use codec?
- 0:57And so for these first eight use cases,
- 0:59that's exactly what we're going to be
- 1:00talking about. And so the first way that
- 1:02I use claude code that I think everyone
- 1:05should 100% be using is Claude projects.
- 1:09And so I made a video about a week ago
- 1:11where I said that you should not be
- 1:12using normal Claude and that you should
- 1:14be using Claude code because it's so
- 1:16much more powerful and it has access to
- 1:19all of the files on your computer. And
- 1:22running Cloud Code locally on your
- 1:23computer allows you to use prompts like
- 1:25this. Hey, there's a folder in my
- 1:27downloads called October 26th vlog.
- 1:29Please edit the video and post it on
- 1:31YouTube. And Claude Code can end to end
- 1:34complete this full task. But there's
- 1:36also another setting in Claude Code that
- 1:38you should 100% be aware of, right? Not
- 1:41only can you run it on your computer,
- 1:43you can also run it in a cloud session.
- 1:46It'll actually spin up a little
- 1:47computer, do anything that Cloud Code
- 1:49might want to do. it's just not running
- 1:51locally on your computer. And for a lot
- 1:53of tasks, that's really useful because
- 1:55you don't actually need your computer
- 1:56for all of the different tasks. And so
- 1:58if you think this is confusing, that's
- 2:02perfectly okay because cloud projects
- 2:04will actually manage this entire
- 2:06process. On the left here, you can see
- 2:08all of the cloud projects that I've
- 2:10created. And here I've opened up editor.
- 2:13And so this main session is just a
- 2:16coordinator session. and it will spin up
- 2:18different clawed sessions whether it is
- 2:21a local session or a session on my
- 2:24computer. So I can say something like
- 2:26this. Hello, I need you to please look
- 2:28at my latest videos in my downloads and
- 2:31I want you to make three clips and edit
- 2:33these locally on my computer and upload
- 2:35them to drive and give me a link to all
- 2:36three. So I can fire that prompt and
- 2:38look at this. It was actually sent to an
- 2:41existing thread. So sent to a thread.
- 2:43Here's a vertical clips from side
- 2:45channel video. And if you see right
- 2:47here, we can see that this is running
- 2:49locally on my computer. And so if I
- 2:52click on threads, I can actually see all
- 2:54of the different threads that are
- 2:56running right now. Some of which are on
- 2:58the computer and some of which are in
- 3:01the cloud. Like this one is in the cloud
- 3:03and all of these threads are are running
- 3:05in the cloud. And so right now we see
- 3:08that my local session is working. And if
- 3:10I click on this, we can see that the
- 3:12prompt that I entered here was
- 3:14automatically passed to this local
- 3:16session. And now cloud code is working
- 3:18and it's working locally on my computer.
- 3:20But I could say, hey, while that's
- 3:22working, I actually want you to look up
- 3:24best practices and create a little
- 3:26artifact for this. Like I want to create
- 3:28like a little webpage artifact that I
- 3:30can use and send to my team on best
- 3:32practices for clipping long- form videos
- 3:34into short form. This can just be done
- 3:36in the cloud. And so I can fire this
- 3:38prompt in the main thread or the
- 3:41coordinator thread called editor which
- 3:43is a cloud project. And here look at
- 3:45this. It says right clip. So it just
- 3:48created a new session. Notice here
- 3:50there's no little computer icon. This
- 3:52one's actually running in the cloud. And
- 3:54this is really useful because I can do
- 3:57all of this from my phone. And so if we
- 3:59open up the clawed iOS app and I click
- 4:02on this side panel right here and then I
- 4:05can see all of my projects right here.
- 4:07So if I click on editor I can see the
- 4:09sessions that are running right now. And
- 4:11so as long as my computer is on, I can
- 4:15fully control my computer using claw
- 4:18projects. Right here we can see that the
- 4:20web one is done. So let me click on
- 4:23this. It should have sent me a page.
- 4:26Here is the artifact that it sent me.
- 4:28And I can very easily see it. This is a
- 4:30sharable artifact that I can share with
- 4:32my team. And this is pretty cool. It's
- 4:35pretty easy. It shows the framing and
- 4:37safe zones, which is pretty cool. And
- 4:40one of my favorite things that I do with
- 4:41Claude projects is I will come down here
- 4:44to the bottom. I will press this button,
- 4:47open up the artifact, and then I'll just
- 4:49speak out the changes that I want to
- 4:50make. I actually want this to be better
- 4:53format and and just make the design a
- 4:55lot better and the top table. change it,
- 4:58please. And it allows me to just record
- 5:00my voice and I can create artifacts. And
- 5:02that's exactly how I created this
- 5:05presentation right here. I created this
- 5:07presentation this morning using Claude
- 5:09Projects from my phone walking back from
- 5:12the gym. I had this idea for this video.
- 5:14I said, "Create this slide deck and I
- 5:17have a very specific skill that I use
- 5:18for it." And when I got back, it was
- 5:20done. And so I use cloud projects to
- 5:23coordinate my cloud sessions, whether
- 5:26that's cloud sessions or my local
- 5:28sessions that are running on my
- 5:29computer. And this is very similar to
- 5:31how I use codecs or their new dots
- 5:33feature. You can very similarly you can
- 5:36speak to your dot and you can get it to
- 5:38coordinate codec sessions. And we'll
- 5:40talk about that a little bit later. But
- 5:42now for the next three items on this
- 5:44list, I want to talk about how I use
- 5:47cloud code or why I use cloud code and
- 5:49what sets cloud code apart from codecs.
- 5:53And so the first one is building apps.
- 5:56Some of you will recognize this. I
- 5:58talked about this in a previous video,
- 5:59but I have my own internal tool that is
- 6:01a mixture of like 10 different SAS apps
- 6:04that I use for everything. And that
- 6:05includes Excaladraw, right? We have all
- 6:07of my Excaladraw documents. This is my
- 6:10own clone of Excaladraw. All of it is
- 6:12stored in my own internal tool. I also
- 6:14have my mind built in which stores all
- 6:16of my bookmarks. I have all of my
- 6:19YouTube analytics and I even have my own
- 6:22video editing platform that I will talk
- 6:24about a little bit later. And in order
- 6:26to manage this entire site, I have this
- 6:29thing called native note. So this is
- 6:31called native note and this is where I
- 6:33build this app and make changes to this
- 6:35app. And so on my phone, I could very
- 6:38easily go to projects and we can move to
- 6:40my agent, my native note project and
- 6:44say, "Hey, what I need you to do right
- 6:46now is I actually don't like how on
- 6:48native note it has my pinned videos. Can
- 6:51you please remove that?" And then also
- 6:53move the conbon title to the top. We
- 6:56also don't need that little subtitle or
- 6:58this like horizontal line. And please
- 7:01make the filter by title thing pinned to
- 7:04the left instead of on the right. And
- 7:06yeah, please make the background have a
- 7:09little bit more texture to it, almost
- 7:10like it's like a white board. And I can
- 7:12fire that prompt off. Notice how I
- 7:14didn't specify whether I want it to be
- 7:16local or in the cloud. It will decide
- 7:18whether it needs my local computer or
- 7:20not. And we've been working a lot. And
- 7:22see here, I've sent the conbon cleanup
- 7:24to this thread, which I can click on.
- 7:27And here it is now building. And here
- 7:30you can see it received the message from
- 7:33the coordinator and it'll begin working.
- 7:35And so you can see here that Opus 5.5 is
- 7:38running on a cloud code session that is
- 7:41running locally on the computer and it
- 7:43is done with three out of the four
- 7:45tasks. It has made all those changes
- 7:48that we just asked and then now it's
- 7:49just waiting to deploy. So it actually
- 7:52shows up in native note. Okay. So now it
- 7:54is done and it should be live. And so if
- 7:57we go to native note, we press commandr
- 7:59and we refresh it, we see that we now
- 8:02have a new app here. It is updated. And
- 8:06so anytime I want to change to the app,
- 8:08whether I'm on my phone or my computer,
- 8:10I can just ask cla projects. All I need
- 8:12to do is go to the native note project,
- 8:14just ask it and it'll figure out how to
- 8:16do it. And so Claude Opus 5.5 right now,
- 8:19in my opinion, is at least for the
- 8:22things that I'm doing, it is the best at
- 8:24coding. And so it's best at creating
- 8:26apps, especially if it is very front-end
- 8:30focused. It's just very good at front
- 8:32end. And GPT 61 soul is also Frontier
- 8:35and it's amazing. It's just not the best
- 8:38at like front-end design. And so that is
- 8:40why I use Claude code specifically
- 8:43projects to build apps. And so the next
- 8:45thing that I use Claude for,
- 8:47specifically Claude projects, is I use
- 8:49it for design and slides. Yesterday,
- 8:52Anthropic released their S1, which is a
- 8:55required document outlining a ton of
- 8:57different things that's required before
- 8:58an IPO. And on my phone using Claude
- 9:01Projects, specifically, I was using my
- 9:03Rick Cra Claude project, which is
- 9:06basically like Claude Duet, which is my
- 9:08own little personal agent, but I just
- 9:10think projects is better. I talked to
- 9:11this agent on my phone. I was like, "Yo,
- 9:13this S1 came out. It's very textheavy
- 9:16and data heavy. I need to properly
- 9:18visualize it, organize it how we
- 9:20normally organize my videos and create a
- 9:22slide deck. And that this is what it
- 9:23created. It shows all of the data in
- 9:26this very easy to understand format
- 9:28where the money goes. I could look at
- 9:30their revenue, their operating costs.
- 9:32And what I've noticed is Enthropics
- 9:34models, especially Opus 5.5, is just
- 9:37incredible at visualizing information.
- 9:40take any information in the world, a PDF
- 9:44that's super data heavy or an Excel
- 9:45sheet, and just tell it to turn it into
- 9:47like this visual thing. Like here it
- 9:49says here is the amount of customers who
- 9:51spend around uh 250 or uh 250 spend 10
- 9:55mil or more. 1,900 people spend a
- 9:58million or more. And it just creates
- 9:59these little visual diagrams here. It
- 10:01shows revenue attention. So a customer
- 10:03spent a dollar a year earlier. That same
- 10:05customer spends $9 now, which is just
- 10:08absolutely unheard of. and it goes
- 10:09through all of its compute. It takes me
- 10:12through everything, all of their risks,
- 10:14and then also it just has really good
- 10:17organization. Many of you who watch my
- 10:19videos know that I have this working
- 10:21outline that gets filled in as we go
- 10:23through the video. And so, it's super
- 10:25cool. And then I had it do a data
- 10:26section at the end. And so, we can look
- 10:28at all the raw numbers. And it created
- 10:30this in three prompts. I just went back
- 10:32and forth on my phone saying like, "Hey,
- 10:34can you change this? Can you change
- 10:35this?" And this is how I learned about
- 10:37everything on the S1 was just with this
- 10:40slide deck. And again, this is not just
- 10:42slide decks too. This is also their
- 10:44claude design mode which is integrated
- 10:46with claude projects. This is actually
- 10:48where I originally designed native note
- 10:51as you can see here. This was how I kind
- 10:53of originally designed the entire app
- 10:55and I just iterated with Claude. I think
- 10:57I did this one on my computer, but you
- 10:59can create these designs on your phone.
- 11:00Anything you do on your computer, you
- 11:02can now do on your phone with cloud
- 11:03projects, which is why I use it so much.
- 11:05Just say, "Hey, create a Figma style
- 11:07design board. Create, you can create as
- 11:09many options as you want and then you
- 11:10can just select the option that you
- 11:12want. You can say, hey, push this to
- 11:13prod or implement these designs into the
- 11:15app that we're creating and it can spin
- 11:17up a local session to do that. It may
- 11:19design it in the cloud and may implement
- 11:21it using your computer and you don't
- 11:23even need to like know that. You can
- 11:25just let your agent do its thing." And
- 11:27so I use it for slide decks and then
- 11:29creating claw design. Now, here's for
- 11:31the most fun one and the most viable use
- 11:34case right now, and that is video
- 11:37editing. And so, I've been using Claude
- 11:39non-stop to edit my videos. In fact,
- 11:42these two videos right here, this how
- 11:44smart people are using AI and how I've
- 11:46been using Claude Duet, both of these
- 11:48videos were fully edited by AI. Now,
- 11:51there was some iteration, but they were
- 11:52edited by AI. And so using Claude
- 11:55projects was also how I've been creating
- 11:57these animated explainer videos with a
- 12:00very specific style. And so as you can
- 12:02see here, I created with Claude and we
- 12:05can actually see which model. So yeah,
- 12:07here are my prompts to create version
- 12:10one of this video. I used these prompts
- 12:12right here and it created this video
- 12:15which explained one of Karpathy's tweets
- 12:18>> with LLMs at all. Around 1 to two
- 12:20billion about 20% are casual and
- 12:23infrequent users of free tier chat GPT
- 12:26like products. This group sees derpy
- 12:28chat bots and treats them a bit like a
- 12:30better Google search or writing aid etc.
- 12:33Maybe an agent tries to book you a
- 12:35flight. I have non- tech friends who
- 12:37reasonably, in my opinion, say they have
- 12:40not much use for it at all. Even many
- 12:42professionals outside of math and code
- 12:44are in this tier. And then that was just
- 12:46version one. This was actually version
- 12:48two of the video. I basically made some
- 12:50changes with these prompts using Claude
- 12:52Opus 5.5 on projects and it made the
- 12:55video much better and it was more zoomed
- 12:57out. And I posted this video on Twitter
- 12:59and it actually got 2,000 likes and 160K
- 13:03views. And many people were commenting
- 13:06asking me about the skills. They're
- 13:08like, "Yo, share how you made it and do
- 13:10you mind sharing the prompt that led to
- 13:12the graphics or is it just Opus 5.5
- 13:14choice?" And by the way, this little app
- 13:16that you're seeing right here, this is
- 13:18V2. This is an app that I'm working on
- 13:21that will be free and open source. It is
- 13:22the GitHub for video editing. And it
- 13:25will allow your agents to create
- 13:28versions of the app that you create. And
- 13:30you'll be able to track all of the
- 13:31prompts and all of the different files
- 13:34that you use to create each video. And
- 13:37people on your team will be able to fork
- 13:39your video. You'll be able to share your
- 13:42videos and make them public. And yeah,
- 13:44I'm really excited about this project.
- 13:46And it is video editor agnostic. You can
- 13:48use any stack, you can use any agent,
- 13:50anything. All you do is copy a prompt,
- 13:52give it to Codeex and it will create
- 13:54version control for your video and all
- 13:57you need to do is go to v2.vide
- 14:00and you can sign up for the wait list.
- 14:02As soon as it's ready, we will send
- 14:04access to everyone who puts their email
- 14:06right here. I'm really excited about
- 14:08this. This is free and this will always
- 14:10be free. This is the GitHub for video
- 14:12editing. Working on it right now with my
- 14:14team. Okay, so in summary with cloud
- 14:15code, it's just incredible at creating
- 14:17things. And I think the best
- 14:19orchestrator of creating things is cloud
- 14:22projects because you can create
- 14:23different projects uh one for each of
- 14:25the apps that you're building. You can
- 14:27create an editor agent and a designer
- 14:28agent and it's really useful. But there
- 14:30are also things that Codeex is better at
- 14:32than Claude Code. And so now we need to
- 14:34move down the stack and we're going to
- 14:36get to codeex. What if your inbox could
- 14:38become a plan you actually use? This
- 14:41part of the video is sponsored by
- 14:43Superhum. As you guys know, I use AI
- 14:45tools for basically everything, and I
- 14:47already use Superhum for email, but
- 14:49pulling together deadlines and feedback
- 14:52still takes work. Superhuman mail and
- 14:54Superhum Docs are part of the Superhum
- 14:58productivity suite. Mail is the email
- 15:00app. Superdocs is a collaborative
- 15:03workspace for documents, tables, and
- 15:05project trackers. Both have connectors
- 15:07for the AI tools you already use. Here's
- 15:10a workflow I want to show you in chat
- 15:12GPT. First, I connect the superhuman
- 15:15mail and ask find the emails about my
- 15:17upcoming videos. What are the
- 15:19deliverable deadlines and latest
- 15:21feedback? Then using superdocs
- 15:23connector, it will turn that into a
- 15:26content calendar with a next action for
- 15:29every video. Flag anything that's
- 15:31missing. Now I have a plan I can review
- 15:33against the actual emails. When
- 15:35something changes, I can ask the AI to
- 15:38update the document. Try superhuman
- 15:40connectors today. You can sign up for
- 15:42superhum using my link in the
- 15:44description. So the first way that I use
- 15:47codecs all like all day every day
- 15:50basically is I use dots with voice. So
- 15:54if we were to measure how I use AI in
- 15:57minutes, this is the number one way I
- 15:59use AI. And I've gotten into a routine.
- 16:02I go to the gym in the morning, but I
- 16:03leave an hour and a half early and I
- 16:06walk around for an hour before I go to
- 16:08the gym talking to my DOT. So, dot is
- 16:11the personal agent platform that is in
- 16:14chat GBT. And I think the correct way to
- 16:16think about dot is it is a codeex
- 16:19orchestrator. The same way that cloud
- 16:21projects can delegate to uh cloud code
- 16:24sessions, dot can delegate to codec
- 16:27sessions. In the same way that cloud
- 16:29code can spin off cloud sessions and
- 16:31sessions running on your computer, dot
- 16:33can do the same thing with codecs. So I
- 16:36can call dot.
- 16:38>> Hey Riley, good to hear you.
- 16:40>> What's on your mind?
- 16:41>> Hey pal, I need you to check the recent
- 16:43changes that I've made to native note.
- 16:45Can you spin up a new codec session?
- 16:47Investigate it. And I actually want you
- 16:49to add that same background that we
- 16:51added to native note to the other pages
- 16:54like to the background of the list view
- 16:56and in the calendar view as well. Can
- 16:59you spin up a new codec session and then
- 17:00uh push your changes to prod
- 17:02>> on it. I'll open a fresh codec session
- 17:04and dig through those recent changes
- 17:05now. I'm getting a new session ready for
- 17:08that background update now including a
- 17:10prod push once it's checked.
- 17:11>> Thank you buddy.
- 17:12>> Of course.
- 17:13>> And look at this. It shows Unifi native
- 17:16note backgrounds. If I click on this
- 17:18right here, we can see that it's spun
- 17:20up.
- 17:20>> Thanks for waiting. The new codec
- 17:22session is running on your MacBook Pro
- 17:24and inspecting native note. The changes
- 17:26and production push are still ahead.
- 17:28I'll keep an eye on it and let you know
- 17:30as soon as there's something to review.
- 17:34>> Okay, that's great. Yep, I can see the
- 17:37codec session running in the sidebar.
- 17:39While that's running, can you actually
- 17:40spin up a new cloud session real quick
- 17:42and just do some research for me on how
- 17:45people are video editing with AI? I want
- 17:48like an in-depth report and you can
- 17:50create a pages, you know, a G in GPT
- 17:53space, you have the new page feature. I
- 17:55want you to create pages. I want
- 17:57diagrams of exactly how this is working
- 17:59and um and uh yeah, that's what I kind
- 18:02of want to know is like what are the
- 18:03strategies people are using? What are
- 18:05the types of skills people are using?
- 18:06put it into GPT pages and yeah, spin
- 18:08this out and put codeex using the GPT6
- 18:11soul model to do this.
- 18:13>> Yep, I got it. I'll start that cloud
- 18:15session with GPT6 soul for the in-depth
- 18:18report on AI video editing with pages
- 18:20and diagrams showing how the tools and
- 18:23skills fit together. And meanwhile, I'll
- 18:25keep an eye on that native note session
- 18:26so we're not doing anything twice.
- 18:28>> Amazing. Yeah. Cool. And then actually
- 18:31once you're done with spinning up that
- 18:32session, can you do you mind just give
- 18:34me a little report here in the chat on
- 18:36like what videos I'm working on in in
- 18:38notion my long form videos?
- 18:39>> Sure. I'll pull that up. I'll check
- 18:41notion for your current long- form
- 18:43videos and put a short status report
- 18:45here in chat.
- 18:46>> Thanks pal. So that's why I use dots
- 18:48with voice. It is like talking to a
- 18:50human that can use codeex for you. And
- 18:53so codeex is like very very good at a
- 18:56lot of things. And the next reason I use
- 19:00dots and codecs as like one of the main
- 19:02tools that I use is it has incredible
- 19:04computer use. In fact, whenever I edit a
- 19:07video and I really like the video and I
- 19:09get to my final version, I just asked my
- 19:11dot. I said, "Hey, the video's ready,
- 19:13post it on YouTube, and what it'll do is
- 19:15it'll analyze the transcript of the
- 19:16video. It will go through in Chrome,
- 19:19upload the YouTube video there, fill out
- 19:21all of the details, get the description
- 19:23ready, look at my emails, you know, if
- 19:26it's a sponsored video and I need to
- 19:28include like a link in the description,
- 19:29it'll add it there automatically and
- 19:31then it will save it as a draft and then
- 19:33my dot will just message me like your
- 19:35dot your draft is created here. Go ahead
- 19:37and look at it and when you're ready,
- 19:38you can post it. It's incredibly useful
- 19:41and computer use with codecs is amazing.
- 19:45Next time you're on your phone and you
- 19:46leave your laptop open, just go to your
- 19:48dot and be like, "Hey, spin up a new
- 19:49codec session and ask it to do anything
- 19:52on your computer and have it take
- 19:53screenshots as it's doing it and you can
- 19:55like kind of see how your agent uses
- 19:58your computer and it's basically like
- 20:00having an agent that can fully control
- 20:02your computer." And so that is one of
- 20:03the main reasons. I also use codecs a
- 20:05lot for business tasks. That's how I
- 20:08kind of split my usage between cloud
- 20:10code and codecs. Cloud code is creating
- 20:12assets for my content business slash
- 20:15when I want to learn anything like in
- 20:16the case of the the anthropic s1 I was
- 20:20just wanted to learn that's why I had it
- 20:21create slides and I do that for
- 20:23basically everything that I want to
- 20:24learn. So if I'm creating an asset, I
- 20:27will basically have Claude code do it
- 20:29because it is just really good at visual
- 20:32design and then also video editing. And
- 20:34codeex I use for my like personal
- 20:37assistant administrator and it keeps
- 20:39track of my calendar, my emails, all the
- 20:41messages in Slack. That's how I use
- 20:43codecs. And I generally just speak with
- 20:46Dot because Dot can spin up codec
- 20:48sessions. So in both of these cases,
- 20:50right, with dots and projects, I'm
- 20:52speaking directly to an orchestrator.
- 20:54And so Anthropic, what I think they're
- 20:56trying to do is they're actually
- 20:58releasing a new product called Claude
- 21:00Duet. Most people don't have access, but
- 21:02it is their version of dots and it is a
- 21:04standalone app. I prefer projects now
- 21:06because Claude Duet cannot spin off
- 21:09local sessions running on your computer
- 21:11and that is a non-negotiable for me. I
- 21:13must be able to do it for me to use that
- 21:15agent frequently. And so that is why I'm
- 21:18using these two apps as my orchestrator.
- 21:21And the final use case that I've been
- 21:23using for Codeex is email and Slack. I
- 21:26mentioned it before. I can't really show
- 21:28examples because all of the data is
- 21:30sensitive. The way I do all of my emails
- 21:32now, and I'm not kidding, like 90% of my
- 21:34emails is I pull out my phone in the
- 21:36morning, put in my AirPods, and go for a
- 21:38walk. and I just talk to the agent and
- 21:40be like, "Okay, you know, every morning
- 21:42there's probably it fil my AI is already
- 21:45set to filter out all of the emails that
- 21:47I don't need to see. It'll just
- 21:48autoarchchive them." And then there's
- 21:50probably like 20 25 emails that I need
- 21:52to hear and I just talk to my dot. I
- 21:54said, "In order of importance, please
- 21:56read me my emails." And then be and then
- 21:58my DOT will talk to me and be like,
- 21:59"This one's from this person and they
- 22:00said this." And then it'll read the
- 22:02email and I'll be like, "Okay, dot I I
- 22:03want you to respond with something like
- 22:05blah blah blah blah blah." And then
- 22:06it'll read it out to me and I'll be send
- 22:08it and then I'll do that on a walk. I
- 22:11love walking. I hate sitting in front of
- 22:13a computer and doing emails. I do all of
- 22:15my email and responding to Slack in
- 22:17little breaks in the day when I go for a
- 22:18walk. When I can go for a walk, I do. In
- 22:21fact, I can't wait to do it after this
- 22:22video and talk to the dot and I can
- 22:24actually get things done while walking
- 22:25around. It's amazing. Okay, so now what
- 22:27I want to talk about is a little bit
- 22:29different of a tool that sits at the
- 22:31center of all of my agent platforms. And
- 22:34this tool is what allows me for my
- 22:37agents to share skills and plugins. And
- 22:40by the way, this is not sponsored at
- 22:42all. It's just an incredibly useful
- 22:45tool. And that is Composeio. And so
- 22:48Compose allows me to centralize my
- 22:52skills and plugins. Many of you guys
- 22:55have seen this screen before in the chat
- 22:57GBT app. And there's an equivalent one
- 22:58in the Claude Code/Cloud app where you
- 23:01can install these plugins and they're
- 23:03really useful because plugins allows
- 23:05your agent to actually use all of these
- 23:07tools. In ClickUp, I store all my brand
- 23:09deals with my management company. In
- 23:11Convex, that's where I store all the
- 23:12data for all my apps. In Gmail, that's
- 23:15where I have all my email. I have my
- 23:17calendar. I have my superhuman email,
- 23:19which I actually use more than the Gmail
- 23:21one. I have my notion where I store a
- 23:23lot of useful things. Versell where I
- 23:25host the apps. Exa where I do it has a
- 23:28better web search than the ones that are
- 23:30built into these agent platforms and
- 23:32then a lot more. And it would be really
- 23:34annoying that if every time I wanted to
- 23:37test a new agent platform, I would have
- 23:39to resign in to all of these. Well, one
- 23:42plugin that you can add to all of your
- 23:44agents is the Composio plugin. And this
- 23:48is Composeio. Once you create your
- 23:49account, all you have to do is go to
- 23:52apps and you have literally every plugin
- 23:54you've ever created. Claude and GPT have
- 23:58no built-in YouTube plugin where you can
- 24:02plug it into your your actual YouTube
- 24:04account where it can change things on
- 24:06your YouTube. They have that on
- 24:08Composio. And so all you have to do is
- 24:11plug in your Composeio and then you get
- 24:12access to all 30 all 30 of these apps
- 24:16that I've added. And you can create your
- 24:18own custom tools like I did here with my
- 24:22native note tool, right? I have an
- 24:23internal tool and I created a native
- 24:26note app on Composeio and all I did is
- 24:30ask my agent to create it. Now all of
- 24:31the agents that I connect to will be
- 24:33able to use native note. It will have
- 24:35all of the instructions and it's really
- 24:37easy to use. And so remember with dot we
- 24:39were asking it to create both of these
- 24:42sessions right here. If we navigate to
- 24:45Composeio and we go to home and we
- 24:47scroll down, we can see that it was
- 24:50actually using the Composeio skills the
- 24:53whole time. The way that it uses Slack
- 24:55is through Composio. The way that it
- 24:57uses Gmail, it's not even using the ones
- 24:59baked into Codeex, it's using the ones
- 25:01on Composeio. And so I'm able to just
- 25:04give my agent the ability to use all of
- 25:07the plugins. And so that means let's say
- 25:09tomorrow Microsoft came out with the
- 25:11best agent platform and it was powered
- 25:13by their new model Microsoft Satia 4.0
- 25:17and this was insane and they had all
- 25:20these different plugins that you could
- 25:21add. All I would need to do is go to
- 25:23this new agent platform and plug in my
- 25:25Composeio and it would have access to
- 25:27all of my plugins. Additionally, they
- 25:29have instant and I'm telling you this
- 25:31feels sponsored. I promise you this is
- 25:33not sponsored. This is just genuinely
- 25:35useful. This is instant. And so what
- 25:38instant does is it basically takes all
- 25:40of these useful API keys that you might
- 25:42want access to like Tavi which is uh
- 25:45search exa which is search. It also has
- 25:48like 11 Labs FAL which has every popular
- 25:52video model for cheaper. It's much
- 25:54cheaper than Higsfield to use. You can
- 25:56just plug it into FAL and get all of the
- 25:59creative video models and image models.
- 26:01What it does is it basically sets this
- 26:02up to a single API. So it'll charge you
- 26:05for the amount that these APIs cost, but
- 26:07you don't actually have to go in and set
- 26:08up your API key. It just automatically
- 26:10connects to all of them. And so the way
- 26:12that I think of Composio, right, it's
- 26:14kind of an agent passport. It allows me
- 26:17to very easily switch from agent to
- 26:19agent platform because I've centralized
- 26:22all my skills and plugins. And by the
- 26:25way, you can also create skills. I think
- 26:27I only showed you the apps that I had
- 26:30built. So in Composeio you have your
- 26:32apps and then you also have skills and
- 26:35then I have mine at my email and here
- 26:36are 20 skills that I use quite
- 26:38frequently. I don't want to have to uh
- 26:40set up skills on Claude code and then
- 26:43tell Claude code to send the skills to
- 26:46codeex or send them to Grockbot. I just
- 26:48centralize them. All of my skills are in
- 26:50one place and all I need to do is
- 26:52connect the Composio integration and it
- 26:55has all of these skills. And the next
- 26:58one is actually one that is something
- 27:01I'm working on right now and I actually
- 27:03don't know how to solve the problem. And
- 27:04for anyone who's solved this problem, I
- 27:06would love for you to like let us know
- 27:08in the comments and that is memory. The
- 27:10same way I centralize my skills in
- 27:12plugins. I want to centralize my memory.
- 27:15Codeex, Claude Code, Grockbot, Muse,
- 27:18Claude Duet, all of these platforms keep
- 27:20a memory on you, but when you switch
- 27:23from platform to platform, the memory
- 27:25does not persist. And so I'm trying to
- 27:27figure out a workflow that is relatively
- 27:29simple that allows these agents to
- 27:31centralize my memory. I tried connecting
- 27:34super memory to Composio and that was
- 27:37okay. I just couldn't figure out how to
- 27:40get it to update its memory more
- 27:42frequently so I could very seamlessly
- 27:43switch from cla to codeex and it could
- 27:47just pick up with all of that context in
- 27:49mind. It would actually have memory of
- 27:51all the things that I'm doing. So, that
- 27:52one's more of an open question because
- 27:54this is the thing that I need to like
- 27:56complete my stack. I need to figure out
- 27:59memory. So, let me know in the comments.
- 28:01Okay. Finally, we need to talk about
- 28:03Grockbot. Now, I don't actually go to
- 28:05the Grockbot UI ever really. I mostly
- 28:08just use Dots and Cloud projects. That's
- 28:11how I interact with agents. But, I want
- 28:13to talk about two things. First of all,
- 28:15Grockbot has Twitter access. And if
- 28:17you're a content creator, it's
- 28:18incredibly useful to get an AI to scour
- 28:21Twitter to find links. Maybe you're
- 28:23looking for B-roll videos to put in your
- 28:25news content. It's really good for doing
- 28:27research and being upto-date, especially
- 28:29on tech news. This is where the smart
- 28:32people are talking about tech period on
- 28:34Twitter. And because you can connect
- 28:35Grockbot to Twitter and then they also
- 28:37have a feature called web hook, you can
- 28:40allow Claude code and codecs to send a
- 28:43request to Grockbot. Let me show you
- 28:46what I mean. So, if we go to Claude and
- 28:48this is a local Claude code session. I
- 28:50can type in Grockbot web hook. Hey, I
- 28:53need you to research Claude design, not
- 28:55Claude design, rather Claude Duet, their
- 28:57new platform. Can you please get
- 28:59Grockbot to do some research on this and
- 29:01create a little report based on this?
- 29:03And so what this is going to do is this
- 29:05is going to use a skill that I've
- 29:07configured called grockbot web hook
- 29:09which is connected to if we open up
- 29:12grockbot
- 29:14we can see here that it has a cloud code
- 29:17web hook skill. So when a web hook fires
- 29:20it will begin working and you can see
- 29:22here that it's automatically starting it
- 29:25is chipping away. It is doing research.
- 29:27It sent a request to Grockbot. So that
- 29:29means I don't need to use three
- 29:31different agent platforms. I can
- 29:33basically just interface with my
- 29:34favorite agent platforms which is codecs
- 29:36or dots and clawed with projects and
- 29:39they can send requests to Grockbot. I
- 29:41don't actually use the Grockbot UI. I
- 29:43just basically have my agents delegate
- 29:45to Grockbot because Grockbot has full
- 29:47access to Twitter. It's really useful.
- 29:49And this also connects to your account.
- 29:51So I could have asked Cloud Code to
- 29:53check my Twitter DMs and it would have
- 29:55just delegated that to Grockbot. Very
- 29:58useful. And there you have it. I am a
- 30:00content creator and I'm starting a
- 30:02company called agent native and these
- 30:04are all of the tools that I use. I use
- 30:06cloud code specifically projects to
- 30:08coordinate all of the assets that I need
- 30:10to create apps, designs, slides and
- 30:12video editing. I use codecs for its
- 30:14computer use. All business and
- 30:17coordination tasks which are just such a
- 30:20high percentage of my business is just
- 30:22coordinating with other people through
- 30:23email, Slack and otherwise. and Codeex
- 30:27can access my text messages as well,
- 30:28which is really useful. That means I can
- 30:30access my messages through voice with
- 30:32dots. And that's kind of the main
- 30:35breakdown. And then obviously plug this
- 30:37into all of the other tools that I use.
- 30:39And the way that I centralize that is
- 30:41through Composio. And then Twitter is
- 30:43such a large part of my business,
- 30:44whether that's doing research or posting
- 30:46on Twitter. It's one of the main
- 30:47accounts. I connect it to Grockbot
- 30:49because it has full access to my Twitter
- 30:51account and doing research on Twitter.
- 30:53So, this is like 90 plus% of the things
- 30:56that I do with my agent stack. And as
- 30:59this evolves, I'll probably remake this
- 31:01video once every two months because
- 31:03it'll probably be different in 2 months
- 31:05because the agent platforms will evolve.
- 31:07Meta might release a new model. Google
- 31:09might release a really good agent
- 31:10platform. It's really, really cool. And
- 31:13who knows, someone might release a
- 31:14better Composio competitor. And so,
- 31:16anyway, thank you guys so much for
- 31:18watching this video. I hope you guys got
- 31:21a lot of value. If there's something
- 31:22that you want me to create or a video
- 31:24you want to see soon, let me know in the
- 31:26comments.
About this transcript
This page contains the full transcript of I Spent $10,000 Testing AI Agents (Here’s the Stack I Actually Use) by Riley Brown, generated from the public captions YouTube serves with the video. The transcript has 6,274 words across 848 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.