스크립트
00:00:00Jack Dorsey and his company Block just released a very exciting new project called Buzz.
00:00:06It's a full-blown open-source workplace platform, very similar to Slack, which aims to change the
00:00:12way humans and AI agents work together in one common workspace. I tested it out myself and
00:00:18it's honestly really interesting. So in today's video, we'll take a look at Buzz, see how it works,
00:00:24and we'll test it out by making a joint human plus AI agent workspace and running some fun
00:00:30experiments on it. It's going to be a lot of fun, so let's dive into it.
00:00:38Okay, so what is Buzz? On the surface, it looks very much like Slack. You've got channels, threads,
00:00:45DMs, media sharing, search. The interface will feel instantly familiar to anyone who's used a
00:00:52modern team chat tool before. But the interesting part is that Buzz is built on Noster, the same
00:00:58decentralized protocol Jack Dorsey has been backing for years. His own Bluetooth mesh app, BitChat,
00:01:05eventually adopted Noster 2 as the fallback transport for where there's no Bluetooth in range. I actually
00:01:11did a video on BitChat one year ago, and fun fact, it was my very first video after joining BetterStack.
00:01:18So go check that out if you're interested. But anyway, the way Noster protocol works is that everything is
00:01:24recorded as a signed event. Messages, reactions, code changes, approvals, all of it lands in one shared
00:01:32searchable audit log. And here's where the agent thing on Buzz gets very clever. Every human and every
00:01:39agent on the platform gets its own cryptographic key pair independent of the platform itself. But the
00:01:46agent's actions also carry a second signature, tying it back to its human owner. So you end up with what
00:01:53Block describes as a cryptographic paper trail that neither the human nor the agent could produce
00:01:59on their own. Basically a verifiable passport for your AI. You can always prove which agent did what
00:02:06and who it was working for. And there are three things that make Buzz different from any other similar
00:02:11platforms. First, it's model and agent agnostic. It works with Cloud Code, OpenAI's codecs, and Block's own
00:02:18Goose framework. It communicates through the agent-client protocol, which is an open standard for wiring coding
00:02:26agents into dev tools. So basically you're not locked into one vendor's assistant. And second, it has Git hosting
00:02:33build directly into it. There's a software forge baked into the app, and every feature branch can become its
00:02:40own channel. So all your patches, your CI results, and your review comments all live in the same place
00:02:46as the conversation that produced them. And that's the part that's aimed squarely at GitHub. And Dorsey
00:02:52And Dorsey described Buzz as a tool to reduce their dependency on Slack and GitHub. And third, you own the infrastructure.
00:03:01It's Apache 2.0 licensed, the source is on GitHub, and you can either run it on your own Noster Relay or use Blocks
00:03:09hosted version, which is currently in beta and is absolutely free to use. So basically you own your data, your relay,
00:03:16and your agents. And currently it's on version 0.4 point something, so it's still very early in its development.
00:03:24So let's go ahead and spin it up and see how it actually works. So the first thing I did was open
00:03:29buzz.x.y.z, just download the app or go to the source and compile it from scratch. Either version worked perfectly
00:03:37fine for me. So the first thing you're presented when you open the app is a button to create your own identity
00:03:43key. And that's the key difference when you're using the Noster protocol. There are no accounts. It's just
00:03:49one key, and that is your identity, which belongs to you and not Buzz. There's no password to reset,
00:03:56and Buzz can't recover your key if you lose it, so keep it in a safe space. Then once you've created your
00:04:01identity key, you'll be presented with an option to connect your agent harness, and then they are added
00:04:07to your path. And then you can also choose your default harness and model combo. And on the next step,
00:04:13you'll be asked to sign in an existing community or create your own. If you create your own, you'll
00:04:19need to log in a relay server. You could technically host your own relay server, but for simplicity's sake
00:04:25of today's demo, I'm going to be using Block's own hosted relay, which does require signing up for.
00:04:32So I went ahead and created an account there, but that is an account with the relay provider,
00:04:38not with Buzz. So that's the distinction. So your identity key is still yours. And once that's done,
00:04:44we can finally create our own community. And for this demo, I'm going to name mine something like
00:04:49Code Dojo. And next, we need to create our profile. For this demo, I'll be using the persona of the
00:04:56famous Silicon Valley character, Richard Hendricks, just for fun. And next, you get this window where
00:05:01we see that we'll be starting out with three AI agents to begin with, Fizz, Honey and Bumble. And
00:05:07now once we launch the community, you can see that we are immediately greeted with these three helper
00:05:12agents in our welcome channel. So right off the bat, you already have a small team of AI agents in your
00:05:18local community ready to help you out with different types of tasks. And as we can see here, each one of
00:05:23them specializes in different fields. And if we want, we can also create our own custom agents. So in my
00:05:29case, I'll create our first agent, which I'll name Bertrand Guilfoyle, keeping with the theme of
00:05:35Silicon Valley. And for the instructions, we're just going to say that you're a skilled web developer
00:05:40focused on writing clean code and paying attention to security practices. And for the harness, I will
00:05:45actually use blocks own goose, because this allows me to hook up external model providers. And in this
00:05:52case, I'll be using open router, and I'll give Guilfoyle the new Gemini 3.6 model. And that's it for our first
00:05:59agent. And once you create your agent, you also get their identity key. So you should always save it in
00:06:05case you need to retrieve it later. So the demo today is going to be a competition between two
00:06:10models to see who can build the best authentication system. So I'm going to go ahead and create another
00:06:17agent and we'll give this one the personality of Dinesh from Silicon Valley. The instructions
00:06:21will be fairly similar. And for the model, I'll give Dinesh GLM 5.2. Next, I'm going to create a new
00:06:29channel called coding challenge. And here I'm going to put these two agents to work in a showdown to
00:06:35create the best auth system application. So they each need to create a separate app without using any
00:06:41third party providers and provide a working demo at the end of the task. And when you see these eyes and
00:06:47comment emojis, that means that the agents have acknowledged your request and started working on
00:06:53it. And you can also see this here down below. And if you click on view activity, you can get a detailed
00:06:59log of what the agent is currently doing. Now, for some reason, Guilfoyle or Gemini Flash, in this case,
00:07:06returned an empty response. So I had to ask Guilfoyle if he understood the task. And this is genuinely
00:07:12hilarious. Look at what kind of response Guilfoyle sent me. Understood, Richard. Building a proper custom
00:07:18auth system with blah, blah, blah. Unlike whatever fragile hack Dinesh is cooking up, mine will actually
00:07:24be secure. I love how Gemini actually applied the spicy Guilfoyle persona for this agent. I love it so
00:07:31much that I'm actually going to react with a joy emoji here. Anyway, while our devs are busy cooking the
00:07:36apps, I'm going to go ahead and create a third agent. This one is going to be Hackerman. And
00:07:42Hackerman is a highly skilled white hat hacker who specializes in cyber security and their expertise
00:07:49is running security audits. And I'll give Hackerman the codex harness with the default model, which in
00:07:55this case is GPT 5.6 Sol. So basically, I want Hackerman to analyze both code bases cooked up by Dinesh and
00:08:01Guilfoyle and determine which one is the most secure one, which in turn will let us know which model is
00:08:07the superior one. In this case, Gemini 3.6 Flash or GLM 5.2. And while I was setting up Hackerman,
00:08:15I see that Guilfoyle has already finished the task. So that was quick and delivered a nice working web
00:08:21app with a cool stylish design. So that's pretty cool. And another interesting thing to note here is that
00:08:27Buzz puts all your projects in the .buzz directory. So that's where all the work done by the agents
00:08:33live. I'm not going to test the whole flow myself. I will let Hackerman be the judge of that. But before
00:08:39that, I need to get Dinesh's app ready. So Dinesh's first result gave me a startup script that failed on
00:08:45my machine. Plus it had two security vulnerabilities. So that's not a good start for GLM 5.2. But nonetheless,
00:08:53I prompted Dinesh to fix these errors. And on the second iteration, we got back a working result.
00:08:59Albeit this one isn't too exciting design-wise, but it looks like it's functional. So that's good.
00:09:05So now I'm going to ask Hackerman to review both of the apps and determine which one is superior in
00:09:11terms of security practices. And a few minutes later, we finally get the result. And to my surprise,
00:09:17Dinesh's app was actually ranked higher than Gilfoil's. Narrowly, but still. So judging by the result,
00:09:25GLM 5.2 beats Gemini 3.6 Flash. Well, at least in terms of security practices. But we also got to
00:09:32take into account the two iterations GLM had to do, plus the vulnerable NPM packages it included. But on
00:09:39the other hand, if we look at Open Router's cost metrics, Gemini Flash is just really, really expensive
00:09:46compared to GLM. So combining all of that, I would say GLM 5.2 comes out being the winner in the end.
00:09:53But now I decided to do something really funny. I asked Dinesh and Gilfoil what are their thoughts
00:09:58about the results. And as soon as I did that, the agents started arguing like crazy. Mainly,
00:10:05Gilfoil and Hackerman had spicier takes and more heated arguments about the results. Dinesh was kind
00:10:11of diplomatic in its answers, but it was honestly so freaking hilarious to see how these agents got in
00:10:16a debate about the results of the test. And another cool thing is that these agents retain a memory of
00:10:22their chats in the community, so they get better over time within the Buzz ecosystem. For example,
00:10:28you might encounter situations where if you're using one of these open router models for your agents,
00:10:33they might end up searching the inner Buzz docs to see how to execute certain commands, like adding
00:10:39others to channels or similar actions. But once they've gone through it, next time they are able to do it
00:10:45in one shot. So overall, I would say that this is a very interesting concept, having a Slack-like
00:10:51workspace for you and your agents. But while testing it, I also came across situations where I exceeded
00:10:57my token limits because the problem with these multi-agent workflows is that they keep messaging
00:11:03each other and piling requests upon each other until you max out your tokens. So what I learned is
00:11:09that you've got to set strict limits for your agents when to do a hard stop on a task, otherwise they can
00:11:15just go off the rails. So there you have it folks, that is Buzz in a nutshell. It's a cool app and it has a
00:11:21really nice design and I had so much fun playing around with these agents in a Slack look-alike
00:11:27interface. But I would say that the app itself is still very early in its infancy. I guess the big
00:11:34pitch here is that this software might become the new place for both Slack-type communication and a
00:11:39single hub for managing Git branches, YAML workflows and even Git hosting. But will people actually latch
00:11:46onto it? I think it's too early to tell. But what do you think about Buzz? Have you tried it? Will you
00:11:52use it? Let us know in the comments section down below. And folks, if you like these types of technical
00:11:57breakdowns, please let me know by smashing that like button underneath the video and also don't forget to
00:12:03subscribe to our channel. This has been Andrus from Betterstack and I will see you in the next videos.