Buzz Didn't Solve What You Think It Solved (Open Source)

Everyone covering Jack Dorsey's new app buzz is telling you it solved multi-agent orchestration. It didn't. I've read the architecture docs. I've read the source code. And now I'm going to break it here live. But what Buzz actually solved is a different problem. And honestly, it might matter a lot more. Quick context. If you don't know about Buzz, Buzz launched July 21st from Block, which is Jack Dorsey's company. The launch tweet alone did 5 12 million views and the repo is at

17,000 stars a week and a half in. It's open- source Apache 2 and the whole thing can run on your server. So here's the mental model for Buzz. So Buzz is one server called a relay and one log. So every single thing that happens, a chat message, an emoji reaction, a git commit, a workflow step, an AI agent that's doing your work is a cryptographically signed event appended to that one log. So you can view it as chat, GitHub, and agents

Collapsed into one system. So it's essentially Slack plus GitHub plus agents on a server that you own. And one detail that tells you how serious this is is the repo was created back in March. They've ran over a hundred releases internally before anyone actually saw it. So, Block essentially built Buzz inside Buzz. All right. Now, I want to go over what happened to the pitch over these six days because this is the part that a lot of people miss. So, day

One here, we have the launch tweet and you can see here he says, "We're building Buzz, a new group chat platform for teams of people and agents built to reduce our dependency on Slack and GitHub." Now, day three, Jack tweeted out here saying, "Every lab will build a social workspace for its own agents. Buzz is the workspace for everyone." That doesn't just sound like a Slack alternative anymore. Next, you can see he tweeted out the open operating system for intelligence and

Action, an operating system. And you can see here someone tweeted out, "This is the first proper multi- aent harness." And he says here, "Whoever wins this layer captures the value as models commoditize." And you can see Dorsy's reply right here is just yes with an exclamation mark. So in less than a week, this pitch went from a cheaper Slack alternative to the OS for all AI agents. And that escalation is exactly what we're going to test today. All right, so let's

Start with what's real here. And the real thing is identity. So in Slack, a bot is a shared token with permission flags. And when something goes wrong, the audit says the bot did it, right? And and you ask which bot, you know, who did this bot do it from? Was it, you know, someone from your team? Was it a client? Whatever the case is, you really have no way to track. And if you've set up any sort of Slack AI bots

And, you know, done anything with that, you'll know it's a huge hassle to set everything up with the proper configs and whatever the case may be. As you can see, we have our own Slack channel with multiple different AI team members as well here. Now, in Buzz, it's a bit different. So, every As a Asian gets its own cryptographic key pair, and it joins channels like a person actually would. So, every action it takes is signed under its own identity forever

In the log. So, the read me here has one of my favorite sentences, which is the same affordances as a human teammate, the same audit trail, a different key pair. And it goes even deeper. So, there's a thing in the code called owner attestation. So every event an agent publishes also carries a signature chaining back to its human owner's teeth. So what that means is the log doesn't just prove which agent acted, it also proves which human authorized that agent to

Exist. And that means an agent can't forge that. Now you may be asking why does this matter? Well, because every company that's about to put agents to work is about to hit that same question. Who did what and on whose behalf? Now Buzz made identity the foundation of this entire tool. It's not just a feature. So to get started with Buzz, you're simply just going to come to buzz.xyz, which is their website, and click on get the app right here. This

Going to download the app. You can open it, install it, and then it's going to look something like this. We'll quickly go over the app and then we'll do some tests right here. So I'll quickly go over some of the features. So when you actually first open the app, you're going to need to create a community. So you can add multiple different communities. You can create a new community or join an existing community. So, if you create a new community right

Here, you're going to be signing into Builder Lab, which is Buzz's hosted communities right here. I'm specifically going to be mainly talking about the cloud version that Buzz has right here. But you can self-host your own community, which I do also have right here. You can see local dev. So, I'll quickly show you how to get both um set up. For the cloud version, it's very simple, straightforward. You just go through the steps for the local version. So you'll see here

In this Google doc right here, which is a free resource that you'll be able to get from our free school community right here. You can access it and you'll see that there is a prompt right here that you can simply just give to one of your AI agents, whether that be cloud code, codeex, whatever the case is. And you'll see this is install and run a self-hosted relay. You can simply just give it this entire prompt and it will set everything

Up for you. If you have any issues, just ask it questions, give it any, you know, errors that it gets and it will get this whole thing set up and to end for you. And then when you're creating a community, you're just going to join an existing community and you're going to paste in this right here, WS or websocket localhost, and then whatever port your um community is on locally. Mine is on 3030, but by default, it's typically on 3000. I'm

Just using port 3000 at the moment. And it would just join community. and then you have it locally set up. Now, if you are using it locally, uh it would work it would be okay for you. If you were using it with a team, they would either have to be on the same network as you or you could use something like Tailscale or if you're really wanting to use this in a production environment, I would recommend setting it up on a

VPS. And once you're actually set up with your Buzz instance, there's a few different things that I want to cover. And we'll first go by starting at the settings area here, hosted communities. So you can sign into builder lab here. Templates. So this allows you to create channel templates so that when you're actually creating multiple different channels and let's say you want to have templates where you can simply just use that template and uh x amount of agents will automatically be

Added to that channel. It can save you a lot of time. So you can set some of these different things up here. Invites. This is where you can invite your team members to your actual community. Agents. Now this is where things start getting good here. So this is where you can actually connect your run times. Now when you first set up Buzz, it's going to detect, you know, what harnesses you're using on your system, whether that be cla etc. And it's

Actually going to install um their ACP, which is agent client protocol. And that's going to allow um you to basically switch with many different uh run times here. It's going to automatically detect specific ones. You can add additional run times, and it will all be connected with buzz. You can see here there's additional run times like AMP, cursor, grock, build, oh my pi, open claw. Uh, and then you can add custom harnesses as well. You set your default harness, which I

Have clawed code here, as well as your default model if you want one. So, Dorsy's favorite feature is this one-click shared compute. So, it essentially allows you to pull your machine's GPU with your team, and agents automatically pick the best available model. It's also only available on Linux and Mac right now. And it's a compile time flag. So, if you're on Windows, the panel loads and then it's going to tell you something like this where it says this feature isn't in

Your build. That's where the product is at right now. Some stuff, some features aren't fully put together or working properly. So, keep that in mind. It is a new repo. Now, here in experiments, this is where you will see the new features that are coming kind of in beta, which I have all these toggled on right here. So, we have workflows, projects, pulse, forms, channels, and agent manage profiles. The workflows, these are like automations and routines. Projects are like the git

Repository and collaboration. Pulse is activity feeds with notes, social posts, and agent activity. Forms is like form style channel threads. And then agent manage profiles. Now, mobile right here is where you can actually go and download the mobile app buzz on the app store or whatever the case is. And once you actually download it, it's very simple. You simply just open the app. is going to allow you to um basically scan the QR code that would be here. You start pairing

And once you actually scan it, you'll be able to actually access it on your phone, which I have it right here, which you can see on the screen here. And you can literally do every single thing that you would normally do on the actual app on your phone. So, first things first, if we go over to agents here, this is where you'll see all your different agents. Okay? When you first start off, you'll have Fizz, Honey, and Bumble here. And you

Can see I also added in a couple different other agents. So I added in Claude Code Fable 5, Claude Code Opus 5, and then we also have Codeex right here, which is GPT 5.6 Soul. I also have another Opus 5 right here, George the Monkey. And then we have Kimmy K3. This one is on the Kimmy code harness right here. And then this one right here is on Buzz's agent harness. If we go to create an agent right here, you'll see

You can discover agents, import agents, or create an agent. So, you would just simply put a name. You can select an image or emoji right here. You can add some agent instructions, which is essentially the system prompt for the agent, or you really don't have to. It's up to you. And then you can either use the harness defaults, which are what we set in the settings here. So, you can see mine is cloud code. Uh you can go to advance here,

And this would allow you to uh have a few additional settings like who can talk to this agent. So, right now, I have it set to only me. If you do allow anyone to talk to the agent, then anyone could uh tag the agent in any channel or you could do a allow whitelist where you could allow specific people in your community to do so. And then parallelism. So this basically is how many conversations each running instance handles at once. So

This is optional in case you want to set a limit for the agent. Now, if you go to customize this agent, this essentially allows you to pretty much connect any specific model that you would want, any specific harness as well within Buzz. So, we have the Buzz agent harness right here where you can select different providers like Enthropic, OpenAI, Open AI compatible, where you could then go ahead and use something like um Open Router or whatever the case is. You can

Use the goose harness, which is another buzz harness they have here. You could use Claude code right here with the Claude models codeex. You could use Hermes, Kimmy, whatever the case is. All right. So, I want to cover a few different things here. So, first things first is if I go ahead and use this prompt right here. So, I'm adding Honey and I say research top the top three complaints developers have about Slack and post a five bullet summary here. When

You finish, mention me, not another agent. All right. So, sometimes I run into bugs where, you know, the agent has to be reset or whatever the case is. So, I'll just go over here. I'll click on the agent and then I'll restart the agent. Sometimes that works. Sometimes I just honestly ask Claude Code, why is this not working? And it will typically help me. But you can see here now we have when it does start working, it will do this emoji

Right here reaction with the eyeballs as well as this conversation one. So it's like it sees it and then it's actually starting the conversation. You can go over here. You see what Honey is actually doing. So it's searching, thinking. We can click on this activity and we can see the commands, the mode, the tokens, what it's searching for, what it's thinking, any tool codes it's doing. Pretty much everything right here, which is really nice. You could also see it in raw

Mode if you want. And so you would see every single little thing right here or you could toggle on timestamps as well. And boom, you can see it posted into the channel right here. So if I look at this thread reply, we can see here are the top three developer complaints about Slack in a five bullet summary. So we have notification overload and constant interruptions pricing that scales painfully poor search and knowledge rot API platform churn for builders and then Slack

AI is underwhelming. Now also too you'll see that I asked Fizz to create me a very basic HTML CRM repository and then Bumble to review and grade this out of 10. And this is very basic. I just wanted to show you guys how it actually works when it comes to the GitHub side of things. So, as you can see, it really just went through both the agents looked at it. It gave its its rating. Now, any type of project or repository

Or codebase that you actually build is going to show up in the projects. And this is essentially like the GitHub side of things. You can see how many repositories you have, your pull requests, local issues, the different people, contribution activity, all your repos, all your pull requests. So, you get the gist. This is, you know, the start of something that I think could definitely be um very powerful. I definitely think they're on to something here with the whole repo projects section

Here. Now, here is something I want you to pay attention to when I give it this actual prompt. So, at George and at Cloudcode Opus 5, work together to plan our Q4 content strategy. When you receive a message from the other, acknowledge it, summarize their point, and ask a clarifying uh follow-up question back to them. Keep each other updated until the plan is complete. We're going to go ahead and click on send here. And you can see both of them reacted

Already as well as now they are getting to work. So we can see both their agent activity right here. If we want to look and see what they're actually doing, we can do so. Okay. So I'm not going to read through every single message thread here, but you can see here we have George that messaged Opus right here and it's giving its response. And then uh Opus is going right here and replying to George. And you can see they're just going

Back and forth. And you can see here at the very end, I did get my actual output after they went back and forth a few different times here. So Buzz solved multi- aent orchestration, right? Well, no. Buzz didn't do any of that. The models did. There is nothing in this actual platform that made them stop. There's no turn limit, no loop detection, no task stats, no dependency graph, no completion conditions. I read that actual source and coordination in Buzz is entirely

An emergent behavior from the actual language model. The Opus decided that the plan was finished and then Buzz just carried the message. And that's the difference. When it works, it's the model. And you know, most of the time I've tried it, it's worked, but when it doesn't, whether you're using maybe not the smartest model or maybe there's some different issue, um there's really nothing to catch it. So this dog right here says they do have a cancel command and their own

Docs say that it's unreachable from every product surface. So it needs an exact text match and a mention tag at the same time and no actual button in the product can produce both. The cancel command right here cannot be typed. So why do I say all this? I just say this to be cognizant and be aware that as you're using different models, whether they may be not as smart as Opus or Fable or maybe just based on the prompt that you

Give it, uh, whatever the case may be, if you're using these models and and maybe you want to be conservative with tokens and whatnot, keep that in mind that if you're using API billing or even if you're trying to conserve your uh, subscription plans with no turn limit or rate limiting, this could increase your bill or really just cause you to hit those limits a lot quicker. Now, keep in mind this is a very new repo, so some of these issues

That I'm bringing up here could potentially be fixed or resolved by the time you're actually watching this. So, we're going to use four different agents right now. We'll use Kimmy Code K3 right here. So, can you do some research on AI news and what is trending? And then George, review and cross reference and fact check everything. And then Codeex, you take this info and develop a list of 10 YouTube video ideas and 15 short form ideas. And then Kimmy K3 on

The buzz harness, you make sure everything is top-notch quality. Let's go ahead and click on send. Now, one thing I will say too is that since this is using the codecs and claude code subscriptions on your actual computer, it has access to all those same skills, all those same setup and configs, etc. MCP servers. So keep that in mind when you're actually using Buzz. Here we can see all four agents are actively working now. All right. And you can see here

We are getting the uh actual responses already. All right. And we can see here Kimmy did some research right here on different AI news topics. And then we see um it tagged Codeex. Codeex then tagged George. And then now Codeex is also tagging me right here. And now you can see right here the issue that happened was George actually went quiet. So we had codeex that was trying to talk to uh George and it was pulling him and it was sending

Direct messages but he wasn't sending anything back. And then you can see here that it said George has not returned the request matrix after repeated pulling and a direct signed mention. The thing is George didn't actually crash. He's mid turn and every time someone actually pinged him for an update, Buzz treated it as a steer and it renewed the deadline by another two hours. It took about 20ome minutes with no answer, no timeout, no error. So essentially, the system cannot tell

The difference between an agent that is working and an agent that is stuck. There's no task state, no dependency, no timeout, nothing reassigns the work. I ended up still getting asking the agent for the output and got the result here. But uh you know this is just one another demonstration I wanted to show you with some of the different limitations with how these agents may work together. So why doesn't Buzz just add a turn limit? Well, honestly maybe they're right not

To. Hard coding how agents coordinate hasn't worked well for anyone. And betting on the model instead of the framework has been the winning bet for 3 years. and they clearly agree because they already defined four rate limit tiers, named them, wrote the config, and then their words, none are actually enforced. So, this is the entire safety system right here, the stop button. And it's their sentence, not mine. And they basically say this requires the user to recognize the loop and know

Where this actual kill switch is. But I'm going to go ahead and send this prompt here tagging George saying, "Set up a workflow for every morning at 9:00 a.m. for you to do AI research online and synthesize this into a morning brief that you send to me." We can see here George is actually working on this uh workflow right now. Okay, so this part I actually found by accident. And this is where I asked George basically to create this workflow like

You just saw here. You can see that he said that he's done and that we have our very own buzz workflow with our own ID and everything like that. But if we go to the workflows tab here, you'll see that there is no workflows. So I refreshed this. I've restarted the app. I checked both of the communities and we have absolutely nothing. So I built the Buzz CLI from source and asked the relay directly and there it is. You can see

The workflow is right here in this channel. You can see the cron and it is signed by George's key. So George created a scheduled job inside the workspace and it's on the actual server. It's signed. It's going to fire every morning, but I can't actually see it in the desktop app. So I can't edit it, delete it, whatever the case is. And the only reason I know it exists is because the agent told me it did. The question you may be

Wondering is, should you actually use Buzz? I would actually say that you should go ahead and test this out. Use it. Try it. Uh, I think this has a lot of potential. I think this really is the way of the future. This is the direction that we're going. This is actually something I've personally been working on within one of my projects. Something kind of similar to Buzz. And honestly, it gave me some good ideas as well. But it's this concept. I

Don't think Buzz is necessarily, you know, we've heard of this concept before. or we've seen it kind of with things like with OpenClaw, Hermes, being able to connect to different messaging platforms, whether it be Telegram, Slack, Discord, whatever the case may be. But these platforms aren't natively built around AI agents. So, this is a na AI native platform kind of like for Slack, for GitHub, or at least with those specific features within it. And I think the future of AI agents

Is multiplayer. And we know this for a fact because we've seen products like claude tag which is so popular internally with with anthropic but I think the power between something like this is one obviously there's a lot of benefits with an AI native uh actual application opposed from something like slack but also it's model and harness agnostic so you can use it with any model any harness. Now, it is very new and there are a lot of different things that I

Would personally improve or want to see within this app, but I think it has a lot of potential. And if you look at the actual team behind it, Block, Jack, Dorsy, you can really see that this, you know, project has a lot of potential. So, other than that guys, I wanted to kind of give you a uh in-depth guide as to actually how to use this, but also a few different unique perspectives that I haven't seen other people talk about. But

Other than that, guys, let me know what your thoughts are in the comments down below about Buzz. Now, also too, I did put together a pretty in-depth uh 20page doc with pretty much everything you need to know about Buzz. A lot of stuff I didn't even talk about in this video, but um just more nitty-gritty details and whatnot. So, you can access this entire document 100% for free. Get access to it in our free school community, the Stride AI Academy. So,

Link to that will be in the description down below, guys. And I'm going to be making a lot more videos in the future surrounding different AI harnesses, surrounding Buzz, multiplayer, AI harnesses, and different tools similar to this, as well as just honestly anything surrounding AI. And I know I haven't been uploading too much lately, but I have been working on a lot of cool stuff that I'm excited to share with you guys very soon. Other than that, guys, I hope you

Enjoyed this video, got some value from it. Let me know what your thoughts are in the comments down below about Buzz. Does it live up to all the buzz that's going around online right now? and let me know if there's something that I missed or something that you think other people would get value from when actually using Buzz, specific use cases that you found or maybe some bugs or other limitations that you found surrounding some of the different areas within Buzz.

But like I said, it's very new and there's going to be a lot more updates coming soon surrounding this. So, make sure to subscribe to stay up to date with that and I will see you guys in the free school community. Other than that guys, I will see you in the next video. Keep hustling, keep grinding and of course guys, accelerate your stride. Take care.