Buzz: The Free Slack Where Claude Code, Codex, and Grok Are Teammates
An interview with an early power-user on why the open-source Slack-for-agents app is going viral, then a full walkthrough of running a personal AI agent team inside it.
Posted
2 days ago
Duration
Format
Interview
educational
Views
31.1K
820 likes
57 · 43
Big Idea
The argument in one line.
A free, open-source Slack alternative called Buzz lets multiple AI coding agents share one conversation and one pool of context, so switching which model or tool handles a task no longer means starting that agent's memory over from zero.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
You already pay for two or more AI coding subscriptions (Claude Code, Codex, Cursor, etc.) and want them working from shared context instead of separate silos.
You run solo or with a small team and want AI agents to draft, critique, and revise work like landing pages or thumbnails without you relaying messages between tools by hand.
You're curious how an open protocol lets different AI harnesses join the same chat, instead of being locked to one vendor's ecosystem.
You want a concrete, step-by-step look at setting up agent-to-agent workflows, not just a feature list.
SKIP IF…
You use exactly one AI coding tool and have no interest in running several models side by side.
You need a stable, production-ready platform today — the video itself flags the automation/workflow feature as buggy and unreliable.
TL;DR
The full version, fast.
Buzz is a free, open-source Slack clone built on an open 'agent client protocol' that lets Claude Code, Codex, Cursor, and other CLI-based agents join the same channels and share context, so switching which model handles a task doesn't lose its history. The host and his guest demonstrate creating agents and channels by talking to a default assistant, running multi-agent brainstorm threads, a voice Huddle feature, a mobile app for remote control, and compute-sharing that lets community members run agents on a shared local machine. In the second half, the host walks through his own working setup: one 'lead' agent holds all his skills and memory, while other agents get pulled in for review and iteration, demonstrated live by having four agents rank and revise a batch of video thumbnails. Scheduled automations are flagged as the platform's weakest, least reliable feature, with a proposed fix of a dedicated verification agent.
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.
The host previews the platform as a free, open-source Slack alternative for AI agent teams and outlines the two-part structure: an interview about why it's going viral, then his own hands-on setup.
01:59 – 07:08
02 · Why is it going viral?
The guest explains that the real draw isn't cloning Slack — it's being an open 'context harvester' that lets you swap models and harnesses without losing shared context, illustrated by a Codex-and-Claude-Code collaboration that researched, designed, and deployed a landing page together.
07:08 – 20:03
03 · Creating agents & channels with the default assistant
The pair create a new research agent and its own channel just by typing a request to the platform's default assistant, get multiple tagged agents to hold a genuine back-and-forth debate, and test the Huddle voice-call feature alongside per-agent parallelism settings for running one harness as several simultaneous sessions.
20:03 – 21:25
04 · Adding other harnesses
Any CLI coding tool built on the same open agent protocol can join the platform, including Cursor and, in principle, OpenCode — the guest suggests simply asking an existing agent to configure a new one for you.
21:25 – 22:36
05 · The mobile app
The free iOS app mirrors desktop agent sessions after a single QR-code pairing, letting the host mention and control his agents from his phone.
22:36 – 25:36
06 · Payments & shared compute
A 'share compute' setting lets other community members' agents run local or open-source models on your machine, which the guest frames as an early step toward a pay-per-task agent economy as frontier-model costs rise.
25:36 – 30:14
07 · Workflows & Projects: a GitHub killer?
The Experiments tab's scheduled-workflow feature is called out as unreliable and buggy, but the Projects feature is framed more ambitiously — pushing and versioning code to the platform's own hosted server as a possible alternative to GitHub.
30:14 – 35:13
08 · The future: an open cowork app?
The host floats his hope that the platform becomes an open, multi-agent version of a knowledge-work app with an in-chat canvas for documents, slides, and spreadsheets, and — since it's open source — talks through forking it himself to prototype that feature.
35:13 – 42:32
09 · Automations, triggers & trusting your agents
Both describe the same failure mode: a scheduled prompt that reports the task instead of doing it, and land on a fix — a dedicated 'task checker' agent that every recurring task must ping to confirm completion.
42:32 – 43:03
10 · Tutorial and breakdown intro
The host closes part one and opens part two, promising a hands-on walkthrough of exactly how he personally sets up his agent team.
43:03 – 43:35
11 · Downloading Buzz and default agents
The platform is free to download and free to use as long as you already have a paid agent subscription connected to it; new accounts start with a default harness and a couple of preloaded agents.
43:35 – 45:13
12 · Creating your first agent
The host walks through the new-agent screen, picking a harness and, for a general-purpose connector agent, choosing an exact underlying model with a set thinking mode and context size.
45:13 – 45:43
13 · Adding agents to channels
New agents don't automatically join every channel — you add them by searching, by @mentioning them once, or by inviting them directly.
45:43 – 47:06
14 · Why Codex is the lead agent
The host explains his default pattern: one 'lead' agent holds all roughly 30 of his skills, notes, and memory and is always his first stop, with other harnesses looped in only for a second opinion or a specific strength like computer use.
47:06 – 48:52
15 · Using three agents on one task
The host fires one prompt that has his lead agent generate a first batch of video thumbnails using a saved skill, then tags two other agents to critique and rank them before a visibly improved second round is produced.
48:52 – 51:23
16 · Adding a custom model via OpenRouter
Adding a model outside the default list requires creating a general-purpose connector agent with an OpenAI-compatible provider, an OpenRouter API key, and a specific base-URL string pasted into advanced settings — after which the round-two thumbnails come back visibly stronger, including one where the agents unprompted inserted the guest's face.
51:23 – 51:50
17 · The iOS app is free and easy to use
The host reiterates that everything shown — reviewing the thumbnail rounds, leaving more feedback, firing off new prompts — works identically from the phone app while away from the laptop.
51:50 – 55:15
18 · The management channel
The host's favorite use case: a narrow 'management' agent that reads his email, chat, sponsorship messages, and texts every three hours and hands back a prioritized action list, which he plans to extend to his six-to-seven-person team next.
Atomic Insights
Lines worth screenshotting.
Buzz is a free, open-source Slack alternative where AI agents like Claude Code, Codex, and Cursor act as teammates inside the same channels, not bolt-on features.
Buzz runs on an open 'agent client protocol,' so it works by launching the underlying CLI tools in a terminal and relaying their output — a Codex session started in Buzz shows up in the regular ChatGPT app's recent chats.
Switching an agent's underlying model or harness carries its full prior context over automatically, because the context lives in the channel's history, not inside the model.
New agents and channels can be created just by typing a request to Buzz's default assistant, with no settings menu required.
Each agent can be pinned to a specific model, so a cheaper or faster model can handle routine summaries while a frontier model is reserved for planning and complex coding.
Tagging multiple agents in one thread and telling them to take turns and not be afraid to disagree produces a genuine structured debate — one user described a 30-minute agent negotiation over contract clauses run this way.
An 'advanced parallelism' setting lets a single connected harness spin up multiple simultaneous sessions, simulating a multi-agent conversation even with only one tool connected.
A voice Huddle feature lets you talk to your agents like a video call; agents filter out messages that don't need a reply, and reliability of the spoken response depended heavily on which model was running.
Sharing your machine's idle compute lets other community members' agents run local models on your hardware, previewing an emerging pay-per-task agent economy as frontier-model costs rise.
A 'Projects' feature positions the platform as a potential alternative to GitHub: it can push and version code to its own hosted database instead of, or alongside, GitHub.
Scheduled recurring prompts are the buggiest part of the platform — agents sometimes describe a task instead of executing it, such as replying that it's time to check email instead of actually checking it.
A proposed fix for unreliable automations: create a dedicated 'task checker' agent that every recurring task must ping, which verifies completion and nudges the responsible agent if the task didn't finish.
One power-user's setup uses a single 'lead' agent holding roughly 30 personal skills and all his memory, and he always starts requests with that agent before looping in others for a second opinion.
A live demo showed four different agents ranking a batch of AI-generated thumbnails, then revising a second round based on that peer feedback — an iterative creative loop with no human input between rounds.
Takeaway
One shared context window is what actually makes a multi-agent AI team useful.
WHAT TO LEARN
The value isn't running many AI models at once — it's that they share one conversation's context, so swapping which model does a task doesn't cost you its memory, and reliability gaps like silent automation failures get fixed with a dedicated verification step, not more trust.
02Why is it going viral?
The pitch that resonates isn't 'a free Slack clone' — it's that swapping which model handles a task doesn't reset its memory, because context lives in the shared channel, not in the model.
If your agents already have your notes and skills connected in one tool, a second agent pulled into the same channel can read that context automatically instead of you copying it over by hand.
03Creating agents & channels with the default assistant
A new specialized agent and its own dedicated channel can be created with one plain-language request instead of digging through settings.
Tagging several agents in one thread and explicitly telling them to take turns and disagree produces a genuine structured debate, not a single merged answer.
If you only have one AI subscription, raising that agent's parallelism setting lets it run multiple sessions at once, simulating a multi-agent conversation.
04Adding other harnesses
The agent roster isn't locked to two or three vendors — any coding tool built on the same open connection protocol can join the same chat.
05The mobile app
Remote control of a coding agent from your phone can already work through a single QR-code pairing, with no lengthy account-linking flow.
06Payments & shared compute
Sharing idle local compute with other community members previews a coming pay-per-task market for AI work, not just a flat subscription market.
07Workflows & Projects: a GitHub killer?
Treat any 'recurring workflow' automation feature as immature until you've verified it actually executes — don't assume a scheduled prompt ran correctly just because nothing errored.
A chat platform built for AI agents can double as a lightweight code host, since project work already lives on a server the platform, or you, controls.
08The future: an open cowork app?
Because the platform is open source, you don't have to wait for the vendor to build a feature you want — forking it and prototyping the feature yourself with an AI agent is a real option.
09Automations, triggers & trusting your agents
Scheduled AI tasks can fail silently by describing the task instead of doing it — verify completion rather than trusting that a recurring prompt fired correctly.
A workable fix: create a dedicated verification agent that every scheduled task must report to, so failures get caught and re-nudged automatically instead of going unnoticed.
11Downloading Buzz and default agents
The cost model worth noting: the platform itself is free, but it only works because it rides on AI subscriptions you're already paying for elsewhere.
12Creating your first agent
When adding a less-common harness through a general-purpose connector, you can pick the exact underlying model, thinking mode, and context window rather than accepting a fixed default.
13Adding agents to channels
New agents don't automatically join every channel — @mentioning one once is enough to pull it into that conversation going forward.
14Why Codex is the lead agent
Designating one 'lead' agent that holds all your accumulated skills, notes, and memory, and defaulting to it first, avoids fragmenting your setup across many shallow agents.
15Using three agents on one task
Running the same creative brief through multiple agents for critique-then-revise produced a visibly stronger second pass than one agent working alone.
16Adding a custom model via OpenRouter
Adding a fully custom model via an OpenAI-compatible endpoint is possible, but the exact base-URL string has to be pasted precisely — even the connected AI agents couldn't reliably supply it themselves.
17The iOS app is free and easy to use
An entire multi-agent review-and-revise loop is fully usable from a phone, not just a desktop-only feature.
18The management channel
A narrow 'management' agent that reads your inbox and messages on a fixed schedule and hands back a short prioritized action list turns constant manual triage into periodic review.
Glossary
Terms worth knowing.
Agent client protocol (ACP)
An open standard that connects different AI command-line tools (Claude Code, Codex, Cursor, etc.) to a shared chat interface, so each can post and receive messages as if it were a teammate.
Relay
The server and database that store a community's messages and shared code inside the platform; a community can let it be hosted for them or run it on their own infrastructure.
Harness
The underlying command-line tool — such as Claude Code, Codex, Cursor, or a Grok CLI — that actually runs a given agent's session behind the chat interface.
Huddle
A voice-call feature, modeled on Slack's Huddle, that lets a person speak to their AI agents live instead of typing to them.
Parallelism (agent setting)
A per-agent setting controlling how many tasks that agent's underlying harness can work on at the same time.
OpenAI-compatible connector
A generic agent type that connects to any model through a standard API shape, used to add outside models (via a service like OpenRouter) that aren't in the platform's default list.
Workflow
The platform's recurring or scheduled task feature, shown in the video to be unreliable — sometimes announcing a task instead of completing it.
Project (as a code-hosting feature)
A feature that pushes and version-controls code to a community's own server, positioned as a possible lightweight alternative to GitHub.
“It's a giant context harvester — a place where all your context lives and can be shared between your teammates and your agents.”
defines the whole product's pitch in one line→ IG reel cold open↗ Tweet quote
39:44
“My brain immediately goes to creating an agent called task checker, and its only role is to make sure that the agents that do tasks are done.”
a concrete, transferable fix for a common AI-agent trust problem→ newsletter pull-quote↗ Tweet quote
46:05
“My lead agent is Codex because I use it the most and it has all of my skills.”
crisp statement of the lead-agent pattern the whole second half is built on→ TikTok hook↗ Tweet quote
50:50
“Teammates, not bots — my AI team.”
tight punchline that frames the video's whole thesis→ TikTok hook↗ 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.
17px
metaphoranalogy
Today, we're talking about Buzz, a new free version of Slack that allows you to create teams of AI agents that work with your existing AI agent subscriptions like Codex and Claude Code. It was created by Jack Dorsey, the founder of Twitter, and it's free and open source. And I've been using Buzz to allow all these different agents to collaborate as a team on different tasks that are critical to my business, and I'm blown away by how well this actually works.
And by the end of this conversation, you'll be able to create a team of AI agents that collaborate to do your work. You'll also be able to connect your existing Claude code and Codex accounts, and you'll understand how to use the main features of Buzz, like creating new agents with specific models, how to create custom workflows, and how to use the iOS app that lets you control your team of agents from your phone.
So this video is gonna be all about Buzz, and as you can see here, I can type a message and send it to Claude Code, Grock running on Cursor, Kimmy running on Devin, and Codex at the same time. You can see all of them read it, and all of them are now making a response.
And now Grock responds first because it's a really fast model. Codex responded, and there's Kimmy.
And as you can see here, all of my agents responded inside Buzz.
And so this video that you're about to watch is divided into two parts. In part one, we're gonna talk with Vinny, an expert on Buzz. He made this video that went super viral on x, and we're gonna talk about a bunch of different things, why it's going viral, connecting all of your existing tools.
We're also gonna talk about the mobile app and some other advanced topics, and we're also gonna talk about automations, workflows, and agent reliability. And after this conversation, at the end, it might be ten to fifteen minutes, I'm gonna talk about my workflow.
I'm gonna talk about exactly how to set up these AI agents inside Buzz. I'm gonna show you my Buzz setup, and I'm just going to riff and talk about basically how I've been using Buzz and why I think it's so cool. Let's hop into Buzz.
Vinny, super excited to have you on today. Yeah. So this is Buzz.
And your video, I think it got over a million views, right, on onX? Pretty crazy. Yeah.
Pretty crazy. Why is this platform
going so viral? What is it, and what makes it so interesting to people, do you think? Well, I think, first of all, people are very interested in seeing something that could replace Slack.
People want to replace Slack, but that is, like, a very superficial reason. There's two things. There's you have this thing of context.
Like, Buzz is a giant context harvester and and a place where, like, all your context lives and and can be shared between your teammates and your agents, and your agents just kind of function as teammates.
And in a nutshell, it's the openness, the ability to swap out models,
and have all this context in one central place so you don't have to go between a million different tools. Great explanation. I wanna just make sure that I fully understand, and I wanna make sure the audience understands what this actually means.
Buzz is basically a clone or a somewhat of a clone of Slack. And when you first sign in and it's completely free, you can either choose to create an organization or a team, and then you can add your agents. And the coolest part about this is not only is it free, but you add your existing Claude code and your existing codec setup.
And my a lot of my audience uses codecs. A lot of my audience uses Claude code. And when you use those agents or or those tools, you have skills that you've already created.
What's cool about Buzz is, as you can see here, I said to Codex, if I zoom in here real quick, I said, take the lead on a new educational page for Buzz.
This should be a landing page that is deployed to Vercel, and then I said, consult with ClaudeCode first. And so, basically, Codex said I'm taking the lead on it, and then it asked a few questions to ClaudeCode. ClaudeCode reviewed its idea, came up with a bunch of ideas.
And because Codex has access to my Notion, my Notion has notes for this episode of what I think is interesting about Buzz. And then eventually, it deployed this link to Vercel.
I've never opened this up. So I'm gonna go ahead and open this up, and this is what it created. This is just a demo for something that like, I wanted to show something tangible here.
This is kinda cool. It just kind of created this based on some information that I already have. One thing I want to make clear for the audience here is when I add codecs, it literally shows up in the codecs or the the ChatGPT app.
You could see here that the sessions that I have, like, when I message codecs through Buzz, it shows up here in the recent chats.
And so it basically can can you help me explain this a little bit? Like, I'm I'm trying to wrap my head around this. It uses something under the hood called
agent client or agent connect protocol. I'm not sure exactly what the acronym stands for, but it's a an an open way for them to communicate with the different harnesses. So it's basically just running your you know, it's it's communicating.
All these are CLI tools, so they're running in the terminal. And you are connecting to them, and you're running the codex commands in the terminal, and that's why it'll show up here. So it it, like, injects all of the context that codex might need in order to, like, create a good response, and then that that's what shows up in Buzz.
And they're able to fly out how much historical context it has, but that's, like, the great thing. Like, I was talking about earlier, you just if you have an agent, for example, you made Harry your agent, and Harry is a has a system prompt to do, I don't know, copy editing and marketing stuff.
Now you you don't really like Cloud Opus for that, and you just go into the preferences and you switch it to codex, and you like Soul for that. It works better, you think. It's gonna have all that context that ClaudeCode had in Opus, and it'll just automatically transfer over and and pass that context over to codex.
That's the really cool part is that having the chat sessions in here doesn't matter how often you switch, which models you choose, which harnesses you choose. The context will always get passed over to those those new harnesses. Yeah.
Makes sense. And, yeah, you're right. I did kind of take a straightforward naming approach.
I do have an agent called management,
and it is also powered by Codex. But I do just kinda have, like, a Codex and ClaudeCode agent that I like to have, which kinda has no underlying system prompt. But correct me if I'm wrong here.
An agent that you create, and let's say I created a new agent powered by Codex, it's basically just Codex with a with an added system prompt to it. Is that basically it? Yeah.
That's that's what I get. I mean, obviously, this is a layer on top of the agent, and there's stuff going on in
Buzz where that you know, it's telling it it has its own system prompt so that it knows how to use Buzz and how to leverage the tools within Buzz. But, yeah, essentially, it's just a layer on on top.
Right?
Did that answer your question? Yeah. 100%.
And so Okay. What's cool about what I found to be really cool about Buzz is that you can create channels, you can create agents, and you can do all of this by asking an agent inside Buzz, which is really cool.
And so I can say, Fizz, can you please create a new agent named Harry who does research on the best models and AI tools in the world?
I can specify that I want it to be powered by Codecs and also make a new public channel called research where that agent will exist and put its work.
And you can just, like, fire this prompt off. I don't know who Fizz is. So when you first join Buzz, there's these, like, preloaded agents.
I I don't know who I think they're powered by Codex,
maybe. I I don't know. I actually have no idea.
When you sign up and or you install the app, it'll also it'll already detect which harnesses you have installed on your system. And then if if you have, you know, one or two, you can click install.
And so it will whichever one you install the ACP, this connector for, to speak with the the agents, it will use that one by default. Yeah. Got you.
Okay. Yeah. So whenever I need to, like, configure anything, I don't know why.
I just talked to this yellow guy named Fizz, which is super funny.
Yeah. Yeah. And so yeah.
So right now, I I think we'll we'll let this run-in the background. It should do it. I like how before it creates, like, an agent, it'll just kind of pop up on my screen, and then I can, like, approve it, which is really cool.
But, yeah, it's like a fully agent native Slack. I'm curious. I'm curious.
How have you been using it? I know it's only been out for a week. What would you say your go to workflow is within Buzz?
I mean, the the real the I'm kind of a I do things pretty vanilla. Like, I I keep things pretty simple.
Like you, I don't go crazy with system prompts. I haven't really given agents any specific roles. Like, I'm using the default agents.
I just the first thing I did was pin them to certain models. So I have I took Fizz, and I just gave Fizz Fable because I wanted it to have a powerful model. And I the other one, I think Honey or whatever, I gave it Sonnet because I want one that you know, good for really simple tasks like just reviewing things, giving me summaries.
But if I wanna plan and code and maybe, you know, come up with more complex tasks, then I'll use the better model, the more powerful model. So that's pretty much what I do. The the powerful thing for me is that everything is like you're doing.
Like, you can get the agents to create different kind of workflows for you, start channels, da da da. And then, you know, you go into those channels. But, basically, what I'll do is a channel will kind of be like a feature or an idea or maybe a task.
So So let's say I'm brainstorming something, like, let's say, a new landing page design. I'll start a new channel, and then I'll start a add an agent and start a conversation with the agent there. If I were to be working with other people, add them to the channel.
We would brainstorm. Then we tag the agent like, hey. We like these three ideas.
Go off in parallel and make these three landing page designs, for example. That makes sense. And so you just brought up working with other people.
And this is actually where, like,
I think it starts to get a little confusing because Mhmm. I've just used this as a personal kind of a way for me to interact with agents in a very novel way. Like you, I create threads for certain things, and the agents will just kind of collaborate.
What I love about Buzz, and I think it's just their system prompt and kind of the way they set up Buzz, it's really good at just saying like, hey. Oftentimes, I'll just say, one of you take the lead, figure this out, and then Codex would be like, okay.
I'll take the lead, and then it'll ask a question to Fable. And then they'll kind of go back and forth twice and then start working on it. So that's, like, really fun to work one one human, many agents.
My question is where where it starts to get confusing is how to invite other people. You know, this isn't. It's just like, how do you host this?
Right? When when my computer's off, obviously, my agents won't still exist. Right?
Like, my agents are running they're running locally. They're using my skills that are located on my computer.
How does hosting work, and, like, how would a team use this if you could break it down in, like, simple terms? Sure. So
when you start up, you each get there's a there's a really core concept of openness to Buzz that might be hard for people to kind of wrap their head around. But plainly, I could I can say that in it's it's very different to Slack in that Slack is the company that controls the data.
All your data is going in and out of Slack. Right? They have the servers.
They have the databases. When you write a message, it gets stored on their database. The approach that Buzz takes is that you use your either a hosted relay, it's called.
But if you think of it, it's a server. It's a database. So what but what Block will do or Buzz will do for you is they will spin up one of these servers with a database for you.
So all the information that's happening on your in your instance of Buzz or in your community is going to whichever relay or server that you either can self host or that you let block self host. So that's, like, the main difference. If you let them self host or I'm sorry.
If you let them host that relay for you, it's functioning pretty much the way that Slack would function, but you, uh, have more control over that data. And we don't need to get too deep into that, but then inviting someone is pretty similar. You send them an invite list, and they get added to your community.
And it all functions like Slack would function.
That makes sense. That makes sense. I wanna go back to the prompt that I ran a little bit earlier.
So I basically said, hey. Can you please create an agent named Harry who does research on the best models and AI tools in the world powered by Codex? Also make a new public channel called Research.
And while you were talking, I quickly approved these screens that popped up, which were to create the new agent. And so if we come to the research thing, and I think we can add people, it should have created an agent called Harry, and it has the Codex logo.
I'm typing out this prompt here. So I'm just gonna say, Harry, take the lead. Talk to the agents here.
Come up with the best way to structure automation so that this research channel is the most productive for finding ideas for my podcast agent native.
Take turns discussing ideas. Don't be afraid to disagree.
And you can just tag all of the agents, and then you can just fire it off. And what I really like is that they'll immediately react. And so I think if they've seen it, it'll show the eyes.
If they're actively working and responding, it'll show this. And then if you hover over this, you can actually see all of the agents that are working.
And not only can you see which ones are working, you can actually click and see their activity. So this is Claude code, I believe, and you can just kind of see what each agent is doing. And, yeah, I think this is this is so much fun to kind of interact with all these agents.
Yeah. This is cool. This this is why Riley is the man because he's thinking of cool stuff like this.
I saw you posted this on Twitter. I hadn't even thought of of trying to, you know, like, have them interact with with each other like that. I didn't think it would actually work.
I thought because of the way you have to tag agents for them to even contribute to the conversation, I thought they would not allow that. But that's, like, a super cool way to use these different models. Right?
Like, each agent uses a different harness and maybe a different model, and you might have them be able to kind of just, like, duke it out and come up with really cool ideas that way. Yes. And one thing I figured out
or at least I've tried this before in in many different ways. Like, I tried to hack together my own solution, and one of the main problems is that they would get into these not infinite loops. They would stop at some point, but they would talk for way too long.
And I found that Buzz, however they set up their harness, they usually just do, like, the minimum amount of turns based on my request. So it's, like, very concise. You know, I was drafting up, like, a legal contract.
I actually can't even show it because it is, like, with sensitive information. But I had Fable and then five point six Soul Ultra. They were, like, discussing all of these different clauses while also going through the contract in my email, like, previous contracts in my email, and they went on for, like, thirty minutes, like, discussing all of the different clauses.
And then Codex found something, and it was like, okay. We should do this. And then Claude searched the Internet, searched the old contract that we've signed earlier, and then, like, came up with a counterpoint.
And it was just so fun to see them, and then they settled on this, like, final agreement. And I've just found that Buzz does a really good job of, like, allowing these agents to, like, concisely discuss, and then it'll create the thing, and then the thread just stops, which is cool. Which is not easy to do.
No. That's some singularity
stuff right there. A 100%. And it kinda reminds me of, like do you remember Malt book?
Like, they like Yeah. It's like a more structured, productive Malt book kind of. But, anyway, yeah, here, Harry said, I'll I'm taking the editor role, and it can it mentions the other agents in its response, and it's just kind of working.
And then it's just it's just fun to see. I I've never put this many agents, but they're just having a conversation.
And
Harry I think it's important to mention that, like, some people might not have lots of, you know, unlimited plans for different harnesses and different models and stuff.
And there is an a feature that you can when you create the agent or edit the agent, you can select how many levels of parallelism.
Mhmm. So you can yeah. If you go to agents, you can allow them to work on task multiple tasks at the same time.
So, therefore, if you only have, like if you only have codex, you can allow it to spin up, like, multiple sessions and work in parallel. And you can set that somewhere in the settings, like, say, five max or something like that. Or you wanna let it go crazy, you can do ten, fifteen, 20.
You know? Oh, I see. Customize the agent under advanced.
Parallelism. Parallelism.
Very cool. Oh, so I I have it set to one. Is that Yeah.
So is that I don't know if it if it really honors that. But, for example, if you just like I was saying, you only have codex, you could bump it up to 10, and you could do the same thing that you're the same thing you're showing here with the with that feature on. I see.
Codex. Yeah. That makes sense.
Very cool. Maybe you won't be able to have codex talking to itself, but you could have one feature saying, like, do this research, and then you could say in another channel,
Codex, you know, work on this app or something. Makes sense. That's very cool.
Yeah. What other settings can you do? So you can create an agent.
You can create a channel. Have you tried the Huddl feature? Yes.
And, uh, I didn't know how to get it working until just a minute ago, I can run you through it if you want. Let's do it. So, obviously, for those of you who've used Slack, uh, you've probably have used the Huddl feature, which allows it's basically just like a quick Zoom call.
And I think it's just voice for now, but I think I can just, like, huddle. Yep. Let's see if this works.
So I can huddle with the I think. Like, I need to, like Yeah.
Add them.
I tried this, and it just, like okay. Oh, is it, like, because it's working maybe? Actually, let's just add Codecs and Cloud Code for now.
Yeah. It should be enough. So we have four secret to get it working is that there's that little keyboard button next to the agent button there in the Huddle.
There's a keyboard button. Yeah. Or that start transcript.
Go ahead and start that. Now go to the research Huddle temporary channel underneath there and start talking. Oh,
I see. Yeah. I never pressed this button.
Hello, guys. Can you guys let me know what we've done today? Don't include any sensitive information.
And then stop the transcript by clicking stop. Oh, I see. Okay.
Oh, okay. Answer.
And so they just respond. And what's really cool is I said three different four different messages to them, and they're only responding to the one that would require a response, which is really cool.
They're they can quickly see whether a response is necessary, which is which is, I think, is pretty cool. I wasn't able to get it to work with Fable, maybe because it just took too long, but with Sonnet. So the agent that I had running on Sonnet was able to reply in voice channels.
So, you know, you get a you get a voice talking back to you. So okay. I actually haven't set a model yet.
So if I wanted to create a new agent, could I just ask it to create a new agent powered by Cloud Code, but it's Sonnet? Could I just ask? Probably.
Yeah. I've never tried it, but you could say, yeah, pin the model to Sonnet. So I could just say create can you create a new agent called Sonnet that uses Sonnet five?
Because I do know that Claude code uses Fable out of the box. So Mhmm.
Is that it? Is that how we would do maybe be explicit and just say, like, that uses Sonnet as the only, you know, as the underlying model that it uses always.
Yeah. Yeah.
Okay. And then so, yeah, I guess, let's a lot of our audience, they're, you you know, they're they're token budgeting, not token maxing.
And so Mhmm. You can add other
harnesses to this. I've only added ClaudeCode and Codex. Can I add Cursor to this, for example?
Do you know? I don't know, but anything that uses this agent client protocol can. And I know that open source my guess is now that Cursor has their agents workflows and tools as well.
So probably things like open code for sure. It might not be as easy as click install when you start up Buzz at the first time, but it is possible. And, of course, we've got agents to help us do that.
Yeah. Exactly. I was about to say.
Yeah. Yeah. So, yeah, if you've downloaded OpenCode, my I think you would just sorry to interrupt.
I would I would probably just ask Claude Code. I said, hey. I have OpenCode
downloaded on my computer. Can you please set it up inside Buzz? I'm sure it would probably work.
Yep. It probably would. Yeah.
And so OpenCode, I think, is one of the best ways to get access to all the different models. I've never used it. That would be a nice thing to check out next.
But I know they also have these, like, subscription models, OpenCode, Go, and Zen. Not affiliated with them, but I really like what they're doing. People want choice, and this is kind of the same principle that Buzz is built on.
So, you know, you can choose those models underneath, and I I would assume that you can use those open source and those other models like Kimi with those plans. And then what we should talk about, I think, is incredibly fun and incredibly effective, which is their mobile app. Have you tried the mobile app?
No. Not yet. Yeah.
So when you you can download the Buzz mobile app, and, again, it's free, and it works from my perspective, it works very similarly to the the Codex remote or the in the ChatGPT app. You can very easily connect to Codex through the ChatGPT app. All you do is scan a QR code and boom.
You have it connected. Do you use your phone for, like, coding tasks at all? Definitely.
I have the Cloud Code subscription,
so I can do remote control. And any session that's running, I can access on the on the mobile app, and I really like that. That's a cool feature.
So here's Buzz on my phone, and it took Cool. You know, and, like, I was worried. I was like, oh, no.
Is this gonna take twenty minutes to set up? It's like one QR code, one scan. Now I have it all lined up, and so I can just go into my coding channel, and I can very easily at mention Claude code, and I can say I'm just gonna say, hi, guys.
And they will respond. And it's, like, very they're, like, perfectly in sync, which is really cool.
Mhmm. So, yeah, highly recommend the iOS app. I remember I think it was you and Greg Eisenberg and a few others.
You guys were talking about payments Mhmm. And some implications in the future about how they're gonna implement payments into this.
I know that payments are gonna be really important with agents. I'm curious how you see that kind of taking shape within Buzz or platforms like Buzz. Yeah.
So you're talking about token budgeting or token efficiency, you know. And
if you want to go down to your handle there, it says Riley Brown, and you can go into settings and go to the compute setting on the bottom left under agents. Go to compute.
There is this option there, and you can it says share compute. Share this machine with your relay. When on, other members can run their agents here.
So what it's basically allowing you to do is the local models and the open source models are getting very good, and you can download. It selects one.
It it checks out what the specs on your machine, gives you a model that would run, like, probably the the most powerful model that would run on your machine. And then when you share that on there, you can then allow other people in your community to use that model running on your machine.
So if you you create aid an agent just like you did before, and when you select the harness, instead of selecting Claude code or or Codex, you select Buzz agent. And then there'll be an option to select the shared model that you're that you're using. And what that then allows is other people in your community can at or, you know, mention your agent and get that agent to do work.
And that work that's being done is running on models on your machine. And that right there is already very impressive and very cool, but you see where the option to then maybe charge small fees for for tasks exist. You know, maybe, like, you've got these these frontier models are running on in data centers that have, you know, the the the most powerful GPUs, the most powerful processors that are on the market at the moment.
So even a really powerful local model is gonna take a lot of resources, electricity and a really powerful computer. So you could be maybe the person that invests in that, and you might have community members all over the world who don't usually have access to those kind of resources who might wanna get a task done, and they only wanna pay, you know, per task.
It's like a pay per task model. They could pay you or they could pay the agent or perhaps even agents could pay other agents.
So you have a lower level model that pays a higher level model to do a task for it. So, you know, we're talking about you you're already looking at the agent to agent economy by giving tasks to more multiple agents and telling them to delegate, figure stuff out, and get it done for you, and they're doing a really good job.
But that's because you have access to these plans, and they're probably being subsidized to a certain extent, so the prices will probably go up in the future. And as that stuff can start to get prohibitive, the the expenses get too high, then stuff like this will make it, um, will be a very cool solution.
So that's really interesting. What other features?
I see some things. I've never opened this tab before. All I've done is kind of create agents.
I see experiments. So we have work have you tested any of these? Yeah.
I have. Workflows,
I haven't gotten to work very well for me. In fact, there's something missing there.
There's probably some bugs, but there's also some implementation stuff they could work out. So workflows aren't the greatest, but I can live with that because a lot of the stuff, you know, you know, workflows you can set up, like, kind of recurring tasks. I'm sure they'll work that out.
But projects, I think, is very cool because under the surface, what they're not really talking about as much is that they're also kind of looking to not only be a Slack killer
and this agent orchestrator and manager, but also maybe a GitHub killer. So I'm not sure how familiar everyone in your community is with GitHub. Should we go into that?
Yeah. We should absolutely go into that. You can think of GitHub, at least the way I conceptualize it.
It's almost like a Google Drive for coding tasks. It it it saves the different versions of your code, and all code is is just like a a folder of files, which is called the repository. And
I think that might be enough precursor into where we're going. Yeah. Yeah.
I think you're right. It's it's a place to put your code remotely and but Git itself is a is a tool to you know, because code is something while you work, you might wanna work on many different versions, try out different different features, try out things.
So you can work in different branches and different and and test out things and always come back to, like, the main working code that you have already. So Git is that tool, and GitHub is one of the most popular like, the main default hosting solution to put it somewhere online so that you can collaborate with others or share the code with others.
Sure.
And so you said they were trying to be a GitHub killer. What did you mean by that? So
I can't speak for them. You know, I don't work for for Buzz, but I do believe that they're trying to go after too many tools at once in a way. I think their focus, knowing Jack and his philosophy is their openness is a big part.
So this idea of choice, like you said, I think the future of being able to run and share compute on local models is a big driver behind the behind Buzz. And, basically, because what we talked about earlier, when you start Buzz, that it creates a server and a database for your community that you create.
Well, that is something that Block, the company that creates Buzz, is is taken care of for you. They host that, and they put your data there. But what as a side effect, you have a server and a database.
You can now push your code with Git there instead of GitHub. So I see.
This projects thing, it allows you to work with GitHub still, but you can also push your code to your Relay, what is what it's called, to basically, to the server that is running and storing the data for your community. You can push your code there as well as a as a remote cloud hosted solution.
Cool.
Yeah. That makes sense. And so, like So if you go back to your app, you'll see projects now, a projects tab under inbox on the top left there.
Oh, projects. Yes. And there should be a button maybe when you I don't think I've set up I don't think I've set up the relay.
I don't think I've I've I haven't fully set this up. It's just running locally on my computer. Wouldn't I need to set up the hosted community?
Would I need to do that first? There might be. There
should be it doesn't matter, though, because if it's running locally on your computer, yes, then it might not show that. But what it's doing is it's every time you create or you code or create an app or something, it'll create a folder on your computer and create, like, copies of that folder.
And so you can have, like, three agents working you know, three different ideas on that one app at the same time, and they will be all managed by your agents and by Buzz. And then when you set up most people will go the shared
server relay route. So your code, they'll be able to push the code and save it to the cloud on your Relay, on your data community's database and Relay. Yeah.
Gotcha. That makes sense. And so you're able to just, like, create a bunch of different versions of whatever app you're creating, and the agents can collaborate, and they might not, like, step on each other's feet or, like, mess up someone else's project.
They can Right. They'll do that locally
on your computer, and then they'll have access to you know, with that projects thing, you can tell them upload to GitHub. They might need you to give them
to they might need to install and log you in to GitHub or whatever, or you can just say, yeah, upload to my Relay. Got you. That makes sense.
I I think so I I'm obviously in marketing content and more, like, business operations than, like, technical side. And I I kinda wanna talk a little bit about kind of what I hope this platform turns into, and I'm curious to get your thoughts. So, you know, I spent a year in Silicon Valley, and there's a lot of people right now talking about creating, like, an open co work, right, or an open codex.
You know, codex is this kind of super app that you can interact with an agent. It can do knowledge work for you. And what I'm hoping and I think the coolest part about codex is their in app browser and just kind of that side panel.
So if you were to create a PowerPoint, if you were to create a slide deck, if you were to create a video, if you create an image, it kind of it can open up in this kind of artifact view. And I noticed here that they actually have, like, a little canvas feature. It's very bare bones minimal.
I forget where it is. There's, like, can it's just like a text. Yeah.
It's right there. Yeah. Yeah.
Where is it? Oh, yeah. Canvas.
Second option. And so it opens up a Canvas. And when I saw this, my first instinct was like, oh, it'd be super cool.
And then this is what I I was like, this is probably what an open co work type app would be. It would be like many agents can operate in an interface that many people already understand.
And then, ideally, it would just be able to create any type of document, whether it's like a PDF, a slide deck. Because, like, a lot of people like me who do knowledge work, like, we're creating documents, and I'm trying to outline different things, sending like, that's what a lot of knowledge workers do, spreadsheets.
I think it would be really cool if the agents could create different types of documents and it would just open it up here in the side in Canvas similar to Codecs.
I'm curious about your thoughts there. Someone it's open source software, so someone's gonna build it. Maybe that person will be you, Riley.
Oh,
so it's open source software. So Yeah. I've never contributed to an open source project before.
That sounds scary. But, like, I could, in theory, fork this and just tell Claude I could do it inside Buzz.
Right? I could just say, can you fork Buzz? Buzz to make new features for Buzz.
Yeah. So okay. Please fork Buzz.
Run it locally.
Okay. I'll do that after this episode. Before you do that well, if it works, that'd be cool.
It might be able to find it. But if it doesn't, you just go, Buzz GitHub.
Look for the Buzz GitHub.
Sure. Sure. Sure.
And I can I can see fork it there? This is good. Like, you can always just fire off, like, context as a secondary message.
Mhmm. Search. This is Jack Dorsey's project.
Search Internet. Find the actual project.
I forgot that it's open source. I can just kind of add my own features to it. That'd be kinda cool.
Yeah. You could have your own,
like, custom version. You don't ever have to submit a PR, but if you wanted to try and get, like, kind of, like you said, like, a browser view in there, it might be able to you might probably be able to pull it off with a a Frontier model. Yeah.
Pretty cool. I'll work on that after the after the episode. What else do you think people will find interesting?
Is there anything else that, like, you've been testing that you think is just, like, super cool about Buzz? I think we pretty much hit on it. I think one thing to to note is, like, you're you're thinking ahead.
I think that's right. I think it's I think they've scoped the feature set in the beginning here to let people kind of play around with it and see what people really want. But I think there's a lot of stuff that's gonna come out of this in the future that will make it even more powerful than it is at the moment.
And I would say, like, it's probably for for your audience, it's probably a great thing to try out already. It's not perfect. I don't know if you've run into some kind of, like, rough edges here and there, but I have.
But I'm sure that really trying to push it with, like, workflows and stuff like that. But the the really cool thing is is that because it's built on this open protocol and it's not stuck on, you know, just some one company's controlling all the data and and the shape of everything, it makes it really easy to integrate stuff with.
So what I did well, that was pretty cool. I did it in a in a second tweet with a video that I made where it's just like I'm publishing data from an app that have really strict guardrails. I don't need an agent to do it.
Right? It's like scraping getting information from the x API and stuff. But it'll send it over every day to Buzz.
So because it has this relay that, like, kind I'm kind of in control of and it's an open protocol, it's very easy for me to integrate solutions, my own custom apps that send information over to my Buzz Relay, and I can get Buzz to make this stuff for me.
Right? It does all the coding. It makes it.
And then, boom, in the morning, I have stuff in Buzz that I can then chat with agents about and start working on right away. So this, like, unifies all of your context into one area, into one app so that I'm not going to Twitter, copy and pasting something into ChatGPT, trying to come up with tweet ideas, I don't know, trying to analyze tweets.
It's all like this circular
closed giant context window, and that's what's so cool about it. That is really cool. And I think one thing you touched on earlier is, like I don't know if you used the term automations or webhooks or triggers.
Are there ways that I could get some external thing to, like, trigger a prompt being sent into Buzz?
Because I think that's what a go ahead. Yeah. That's something sorry.
That's something I I tried where it wasn't really working well. So, originally, I was trying to get the agent to, you know, find this information from my app and then kind of summarize it, and it wasn't working.
And maybe you you're pretty good with getting the agents to communicate with each other and stuff and doing it agent first. So that I would maybe try that, but I do think there's potential there.
I think it's a bit buggy at the moment. Oh, wow. So, yeah, it's starting Docker for you.
So there you go. Cool. But, yeah, there is ways.
There are the workflows they're called. So you could say, like, in the in your channel, you know, tell me, ping me every day to yeah.
Let remind me to to do something. I don't know. It it will help solve those word fun.
Yeah. So I could tell Harry, who's my research agent, I said, like, every day, please
search wait. Which model is Harry again?
Oh, no. Harry's Codex. Please, day, search the 25 creators in my niche for AI news and use scrape creators, which is a skill that I've set up.
It's literally scrapecreators.io or something. I forget what it is.
That allows you to just scrape from any social media platform with one API key. And it's like, please find those creators now based oh, Buzz is doing something.
Creators now based on me, Riley Brown, on x.
I don't know. But the point is here is, like, I want to be able to, like, create a new channel. There's things that I do every day, things that I check every day, things that need upkeep every day.
I would love to be able to just tell an agent to, like, handle something every day at a very specific time and then, like, have trust that it solves it. I think that's the one main limitation I'm running into right now as I'm using the all of these different agent tools.
It's like every new task that I create create has, like, this, like, hidden tax on my brain because I'm worried it's not gonna succeed. And, like, the point that I hope we get to is, like, I could just fire up an agent with Fable six and or it's like Fable six on continuous mode, and I could just say, hey.
Like, for these hours, I need you to constantly check on certain things. And, like, here's the end goal every day for this thing and just trust that it'll get done correctly because that will just give me just an insane amount of leverage.
Because right now, like, it does things that surprise me in a very good way, but I just don't have trust in my AI agent team. That's what I lack.
Like, sometimes it doesn't fire off. It'll just, like, some and I've actually tried this already. I don't know if you've seen this where I said, like, every, like, at 9AM, I need you in this management chat, I have it checked my emails.
And then this morning, it just said it literally just said, it's 9AM. I need to do this.
I need to check Riley's email, and it listed out the task instead of doing the task, which is super annoying. I don't know if you've run into That's the problem I've run into with workflows. Exactly.
So they're
I think that's a bug
that they'll probably end up working out. Maybe maybe the the the the current fix might be to, like, whenever you do this task, ping Claude code and then tell Claude code to make sure that it does it so that, like,
Claude code will just do a check. And if it doesn't do the task, it'll tell Codex to finish the task or something. I don't know.
That's what I was trying to do before I hopped on here was, like, get it to I had that app, and it's pushing tweet stats. It's getting all the stats on the my the team I work for, the all of our tweets, the impressions number, the tweets themselves.
And, like, I want push it every day, and automatically, I want it to give me a review of, like, what's the common thread between the stuff that's working, where can we improve, you know, something like that. And it was doing the same thing. And I my idea was to to have it mention another agent and get it to do that, but I haven't found a way to get it working reliably reliably yet.
My brain immediately goes to, like, creating an agent called task checker, and its only role is to, like, make sure that the agents that do tasks are Good idea. Are done.
And then but on every recurring task that you create, the agent that runs will ping this task checker every time no matter what. Maybe that's, like, the first thing it does is ping it, and then the task checker will do nothing if it does it. But if it doesn't do something, it'll be like, hey.
You didn't finish the task or so. I don't know. That's
a good idea. Or just, like, run have it run every fifteen minutes, and if there's a task that didn't really fire, get nudge them or something.
Yeah. Stuff like that. Really good idea.
Yeah. Very interesting.
Yeah, dude. This is incredibly informative.
I'm super excited to dive into workflows and kinda figure that out, and I'm actually gonna fork Buzz. I'm gonna see what I can add on with this Canvas thing.
I think that'd be kinda fun. It's just a side a fun side project. Maybe I'll make a video on that.
I do wanna close. Like, is there anything that like, what are what are you working on? Is there anything you wanna wanna talk about for stuff that you're working on?
I don't know if you're working on any projects or tools and such.
So I'm from what I gather, most of the audience is focused on, like, entrepreneurship and and stuff like that.
But these these guys are great at coding and making apps. And now with these kind of apps that are are very open, you can build stuff that, like, that you can build really bespoke tools that integrate with Buzz and, you know, close this con close the loop on the context that you're kind of, like, shuffling around everywhere.
And I'm part of a team, and we we're working on a full stack TypeScript framework that works really well with agents.
So in my videos, it's called WASP, the framework, which is kind of funny because we're Buzz and WASP. And it does a lot of the framework does a lot of the heavy lifting and gives tools back to the agent so the agent can do stuff very easily.
So, like, when I tell it, use WASP to build an app, like a a CRM, for example, and deploy it right away so that it's live on the Internet. It works really well. And I did that in and I've I've u I'm using Wasp in the videos and showing that.
So that's what I'm working on. We're we're working on a way that, like, makes it very easy for humans and agents to collaborate
to build full stack apps. That's kind of my focus at the moment. Dude, thank you so much.
I really thank you for for kind of talking me through this. I learned a lot, and I'm inspired to kind of work on workflows and also, like, building my own stuff on top of Buzz and kinda see where that takes me.
Yeah. Thank you. Anyway I'm looking forward to seeing what you build.
Alright, guys. You made it to part two of this video, and I hope you enjoyed the conversation. I just wanna make sure that I pack this video with as much value as humanly possible.
And I wanna talk a little bit about how I've constructed my team of these different agents, all running different harnesses. And as you can see here, I got my whole team here.
I have them all responding to all of my messages. They're constantly checking each other's work, and I just wanna take ten to fifteen minutes to describe how I'm thinking about this and how I'm gonna use within my startup.
And so I just wanna start from the beginning. If you haven't downloaded it, you can download it at buzz.xyz. It's literally free to download, and it's free to use as long as you have either Codex or ClaudeCode or any other agent that you've downloaded on your computer, which will actually cost money.
But once you download Buzz, you should come to a page like this, except obviously it'll be empty. And when you're first signing up, you're gonna choose a default agent.
For me, my default agent is Codex. You can use Claude code if you have a Claude subscription as well. And so when you're creating an agent, when you click new agent and create from scratch, and what you can do is you can use your harness defaults, and my default harness is codex.
I can customize this or for this agent, and I can change to any other harness that's running on my computer. Now there's a certain list of tools that will automatically show up if you have them downloaded on your computer.
For instance, Devon is not one of those tools, or at least it wasn't yesterday, so I actually had to ask Codex within Buzz to configure Devon.
Because I have a Devon subscription, I said I need you to configure Devon so it's viewable within the Buzz app, and it can do that. And if you have cursor downloaded on your computer, it should show up here. For instance, I've downloaded Grok build, which is Grok's CLI tool, and I can select Grok Build.
And I've actually never done that. So we can actually configure a Grok Build agent, and we'll just call this Grok Build agent. And so let's say I wanted to create a cursor agent, and you the reason you might wanna create a cursor agent is it allows you to create use of, like, from a very long list of models.
Right? Let's say I wanted to create one with, like, Sonnet five. We want to save money, and we can just do Sonnet five thinking true, context 300 k, and we could just create this cursor agent.
And I could just call this Sonnet, and I could just type in cursor. And so I could choose any image I want.
In this case, I'm going to be using this little quad icon here, and I could just create this Sonnet agent. And as you can see here, I now have added a new agent which is just Cursor which runs Sonnet as the default model. Once you create a new agent, it doesn't automatically enter all of your channels.
You can add them in two ways. One way is you can manually actually three ways.
One way is you could look for them. And so I'm gonna add Muse OpenRouter to this channel.
So this is Meta's new model that I've added. I could also just at mention I could just at mention Sonnet, and I could just say hi.
And when you do this, it should automatically add it to that channel. So all you have to do is mention it. And so the main reason people don't like to switch from codecs to Claude Code, to cursor, to dev, and to all these different harnesses is because usually your skills just live in one place.
And, yeah, you can find ways to, like, merge the skills, but they don't work in the same exact way. And so that's why so far, as I've used Buzz, I will kind of have a lead agent.
And so my lead agent is Codex because I use it the most and it has all of my skills. Another thing that Codex has is computer use. Its computer use is the best, so I will rely on it the most.
So I just kind of use Codex. And so let's say and Codex has all of my skills. And so these are all of the skills that you've already added to the desktop app.
If you watch my videos and you've used Codex, and let's say we're using Codex and we have all my skills here. Right?
I have YouTube researcher skill. I have my thumbnail.
I have all of my different skills. I have many different, like, Notion research skills, probably 30 different skills.
And so Codex really knows me. Codex has all my memory.
I use it the most. And so that's also why I rely on Codex the most. But sometimes you want them to work together, and so I'll make sure to always start off with Codex if I need to use one of my skills.
For instance, I can use Buzz and I can say, like, if we are in our content channel, I can just say, hey, codex, uh, please, I need thumbnails for today's, uh, episode on Buzz.
I need five options. Use sub agents and, um, and make these thumbnails using the correct skill.
And I will just tag in let's go Claude Code and Codex or and Cursor.
And let's say Grock. Maybe Grock has a good sense of what would look good.
Please provide feedback, then codex do another five thumbnails.
And so I can just fire these off and then all of these agents are going to work. As you can see here, we have Claude Code and Codex beginning some activity.
There's Grok. Codex should take the lead. It'll generate some thumbnails.
It's gonna ask Claude Code and Grock for their feedback, and then Codex will do another five, and then I can just include, like, based on their feedback.
But I think that was implied. And now I'll just go I'm gonna go fill up my water, and I'll be back in about five minutes. Okay.
So it used my skill and here's round one. Here's five thumbnails that it created and all of them have my face because it found an image on my computer and it uses it as an input image.
It's very white. So it says please rank these one through five. It tagged Claude Code and Grok five.
And I'll use both critiques to make the next five. And now we can see that Grok and Claude Code are now working, which is really cool to have these agents work together.
While this is loading, I do wanna show you how I added Muse Spark powered by OpenRouter. So Muse is open is Meta's new agentic model that they've created. As you can see here, some feedback just came in from Grock, which is pretty cool.
But if we're going to our agents page here and we wanna create a new agent, create an agent, you can customize this agent and you can create a Buzz agent. And so I think this is their default harness or something.
I actually don't fully know how it works. I just know that you have to do this to use OpenRouter. And the way that I created this one, if we go to edit, we can see that we chose Buzz agent.
We use an LLM provider, which is OpenAI compatible. Right? This is the one that I selected.
And then you need to go to OpenRouter and get an API key. So OpenRouter allows you to use all these different AI models. And so you need to select OpenAI compatible, paste the API key, and then you need to go down to advanced because this model actually won't show up here.
What you need to do is you need to paste this exactly like this. You need to paste I'm gonna copy this.
This is exactly what you need to paste into this right here is underscore base URL. And then you're going to paste this exactly.
If you ask Claude Code or Codex what you need to paste in to use like, your AI agents will tell you exactly what to paste in, they actually can't do this for you. And then for thinking effort, you need to put inherit agent default, inherent agent default, inherent agent default, inherent agent Once you do that, you will be able to click on this model and you'll be able to choose from any model.
And you can create a Buzz agent with any AI model that exists right now. And I did it with Muse, and that's how I added Muse.
So that's just one thing that's pretty interesting. Okay. It looks like it's done.
Round two is done. Five revised. There we go.
Teammates, not bots, my AI team. Oh, look.
There he is. There's Vinny, the guy we interviewed. Right?
They put him in. This changes everything. Look at how it did that.
It started off with these thumbnails right here, very mid. Now Grock and Claude Code gave its feedback, and then look at what it created.
1,200,000 views, one team chat, my AI agent team. This one might be the best one, but it inserted me into all of the thumbnails.
And this is just one way that you can collaborate with AI agents. And remember, on the free iOS app, all of this works. Right?
I can see all of the images that it just created, and I can provide even more feedback and I can say I love five. Let's make three variations and then I'll just at and I can fire this prompt in And now you can see Codex is typing.
And so I can control this from my phone when my laptop is open. Okay. To kinda close out this video, I do wanna talk about one specific channel that I've created.
I can't open it up because it's so personal to me. It manages my email. It manages my Slack.
It manages my communication with my sponsorships team, and it also has access to my text as well. And so I created a new agent powered by Codex. And if you've used Codex, the Codex app on your computer, remember, it's like basically the same thing.
You're just kind of communicating with the same underlying agent, and this agent can access your text messages. And so this management agent, basically, every morning, it reads all of my stuff and then tells me a ordered list of things that I should take action on.
So if it reads an email that I've got that requires, like, an urgent response, it'll put it at the top, and it does this every three hours. So 9AM, 12PM, 3PM, and 6PM.
Every day, it kind of analyzes everything and tells me exactly what I need to do, and these are, like, really urgent things that I need to get done. And I can just respond. I can say, hey.
Can you write this email back to this person? And I can just type it directly in the management chat, which is incredibly useful. And so this management channel has, like, a narrow version of codex, and this management agent has one little paragraph that describes exactly what it does, where it can find key information, the exact link to the part of the Notion that I use.
Right? The these it actually just uses all of my management skills that I already have in Codex. And so when I created this agent, I just asked Codex, I said, hey, there are a lot of things that I use to manage my life.
There's a lot of skills and memory. Can you please take all of that and, like, list them all out here? And then I just told it what parts were important and of those things to include it in the new management agent.
And so now I have this narrow management agent that knows it's a management agent, but it's not a separate agent than codex. Right? It's still the same codex, just a different system prompt that it runs every single time I talk to it, but at least it keep it it keeps it focused, and it just basically lives in my DMs or in the management channel.
Those are the only places I really use this management agent, and I'm now starting to think about how I wanna add teams. And wanna add members of my team. Right?
My content team is, like, six or seven people now. And so I'm starting to think, how do I bring them in to this workspace?
How do I allow them to use all these skills? And these are things that I'm thinking about. And so I'm filming a video tomorrow, actually, with Vishal.
And we are gonna be talking about creating teams, creating agents that don't die within Buzz, and we're gonna dive even deeper. Because whenever I find a tool like Buzz that's and the last time I had this feeling was kind of Open Claw back in January where I'm like, okay.
There's something here. There's something novel about this experience. I kinda wanna follow it deeper, and I I I just I have this instinct that this is, like, a really important form factor for AI agents.
I've never had this feeling of allowing my agents to collaborate in such an easy way that, like, my entire team would understand. It's it requires, like, no technical ability. And so I'm gonna keep using Buzz.
I'm gonna keep learning Buzz, and I will do my best to convey this over the next few videos that I make on Buzz.
I really wanna make it easy to understand and really easy to use. So thank you guys so much for watching. I genuinely am incredibly excited about this this tool.
Um, Vinny was an amazing guest to have on the podcast, and I will see you here for the next video.
The Hook
The bait, then the rug-pull.
The video opens with a promise: a free, open-source Slack clone where Claude Code, Codex, Grok, and other AI coding agents sit in the same channel as teammates, not bolt-on features — and by the end, the viewer will know how to build one.
Frameworks
Named ideas worth stealing.
46:02model
Lead agent + specialist agents
One agent holds all your accumulated skills, notes, and memory and is always the default first stop for any request; other harnesses or models are pulled in only for a second opinion or a specific strength, such as computer use.
Steal forany solo operator running more than one AI subscription who wants consistency without permanently picking a single tool
39:44concept
Task checker reliability pattern
Every recurring or scheduled task pings a dedicated checker agent
The checker verifies the task actually completed
If it didn't, the checker nudges or escalates to another agent to finish it
A proposed fix for AI agents that acknowledge a scheduled task instead of completing it — insert a verification step instead of trusting the schedule to have run correctly.
Steal forany automated or recurring AI workflow (reporting, monitoring, reminders)
47:54concept
Critique-then-revise multi-agent loop
One agent generates a first batch of work
Other agents, running different models, rank and critique it
The original agent revises a second batch using that combined feedback
Demonstrated on video thumbnails: four different models ranked round one, then the lead agent produced a visibly stronger round two based on their combined critique.
Steal forany AI content-generation task where quality benefits from a second opinion before a human sees it, such as thumbnails, copy, or design drafts
CTA Breakdown
How they asked for the click.
VERBAL ASK
43:09link
“It's literally free to download, and it's free to use as long as you have either Codex or ClaudeCode or any other agent that you've downloaded on your computer.”
A soft, reassurance-style CTA placed mid-video at the part-one/part-two transition rather than at the close — positioned to remove the 'is this another subscription' objection right before the hands-on setup walkthrough begins.
Claude Desktop and the Codex app both quietly shipped native, multi-tab in-app browsers within days of each other — used right, one prompt now opens, drafts, and stages a dozen tabs at once.
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.
A 22-minute weekly roundup covering the open-source model that finally passed the vibe check, Codex screen-to-skill recording, the SpaceX/Cursor acquisition, and two Claude updates.
A first-look test of Codex's new talk-back voice mode: triaging a real inbox, editing a whiteboard, building an iOS app, and controlling it all remotely from a phone.
A YouTuber wires Claude Code into an agent-native design canvas called Paper and lets it generate thumbnails, Instagram graphics, and a full pitch deck from plain-English prompts.
00:50
02:08
02:53
03:34
04:14
04:52
06:06
06:34
07:38
09:16
09:58
10:42
11:08
12:21
12:31
13:28
14:09
14:33
15:13
16:13
16:55
17:36
18:18
18:59
19:40
20:45
22:07
23:49
24:10
25:07
26:26
27:22
28:00
28:37
29:10
30:55
31:25
31:50
32:48
33:09
34:11
35:27
36:15
36:56
37:27
38:18
39:00
39:42
40:04
40:50
41:38
44:32
46:37
47:59
49:22
50:04
50:45
52:49
53:31
54:12
54:54
Video of the Day23:10
47:51
22:18
16:27
23:58
21:08