On 21 July, Jack Dorsey’s Block released Buzz, a free and open source workspace that looks a lot like Slack with a code host stapled to it. Channels, threads, direct messages, voice, git repositories, and workflow automation, all in one app, Apache 2.0 licensed, self-hostable, with a managed option if you want Block to run it.
The recap is not the story. Every AI blog has already done the recap. The story is one design decision buried under the launch coverage, and it is a decision we think most of the industry will end up copying.
Agents are members, not integrations
In every tool you already use, an AI agent is a guest. It arrives as a bot user, or an app integration, or a webhook with a name and an avatar. Its identity belongs to whoever installed it. Its history lives in a vendor’s database. When the API key rotates or the workspace moves, the agent is a new thing with no past.
Buzz does the opposite. Every participant, human or agent, gets its own cryptographic keypair. The agent has an identity it owns, a signed history of everything it has said and done, and a reputation that travels with it. Buzz runs on Nostr, which is what makes this possible: every message, code review, and workflow step is a signed event in a shared log. Block’s own framing in the README is blunt about it. Agents are members, not bots.
Sit with the practical version of that for a minute. An agent opens a patch, another agent reviews it, a person approves the merge, and the whole chain is signed by three distinct identities in one log. You can answer “who did this, and can you prove it” without trusting the platform to tell you the truth. Nobody had a good answer to that question before, and it is going to matter more every quarter as more of the work in a business is done by something that is not a person.
The git side follows from the same idea. Buzz uses NIP-34 git events, so patches, repository announcements, and status updates are Nostr events like everything else, and a feature branch can just become a channel: the discussion, the patches, the CI result, and the merge decision all in one place instead of split across a chat app and a code host. Anyone who has spent a morning reconstructing a decision from a Slack thread, a pull request, and somebody’s memory will recognise why that is appealing.
What we actually ran
We have been running Buzz both ways, self-hosted on our own box and as a local desktop client, with Claude Code agents as the workers. Not a demo workspace, real work.
Three things stood out.
The first is that having the conversation and the repository on the same surface changes how you work more than it sounds like it should. You stop context-switching. The decision and the code that came out of it are in the same thread, which means six weeks later the reasoning is still attached to the change.
The second is the audit trail, and this is the one we keep coming back to. When an agent does a piece of work, you can see exactly what it did, signed. Not a log you have to trust, a signature you can check. If you are letting agents touch anything that matters, that is not a nice-to-have.
The third is cost, and it is the least discussed part of the launch. Buzz works with the Claude Code and Codex subscriptions your team already pays for, rather than demanding a per-agent API key and a per-seat licence. Adding a fourth agent to a channel is not a fourth line on an invoice. There is also an experimental shared compute feature where machines in a workspace can contribute capacity and jobs get routed to whatever is available. We would not run that against anything sensitive yet, the isolation and revocation story is not written down clearly enough, but the direction is right.
That combination is the same argument we made in build, buy, or host it yourself. A per-seat product prices you for growth. Buzz prices you for nothing, and hands you the maintenance instead.
The part where we tell you not to switch
Here is the limit, and it is not a small one.
Buzz is early. The version numbers are still in the zero point four range. Block says so themselves, in the README, in plain language: not finished, we will tell you what works and what does not. Their own feature list is split into what works today, what is being wired up, and what is a strong opinion with no code behind it yet.
The pace is genuinely fast. Mobile was on the not-yet list at launch and shipped eight days later, on 29 July, on both iOS and Android, with push notifications built on a protocol that keeps message content and public keys away from Apple and Google. Pairing still needs a desktop app and a QR scan, and the team says they are updating the apps every few days. That is a project moving quickly and telling you the truth about where it is.
Fast is not the same as stable, though, and instability is what we hit. Things move between updates, and something that worked on Tuesday behaves differently on Thursday. That is not a criticism of the team, it is exactly what a project at this stage should look like, and we would rather have honest version numbers than a one point zero that lies. But it does answer the question a business owner is actually asking.
Should you move your team onto this? No. Not this quarter.
The counter-argument is worth naming, because it is a real one. If you are a small technical team with no Slack history to migrate, already running agents daily, and comfortable on your own infrastructure, the calculus is different. You are not migrating, you are choosing a starting point, and the cost of being early is much lower when there is nothing to move. For that team, Buzz today is a defensible bet.
For everyone else, the honest read is that the idea is further ahead than the software. That is a good problem. The reverse, polished software built on a bad idea, is the one you cannot fix.
What this actually signals
We think agent identity is the thing to watch, not Buzz specifically.
Right now most businesses adopting AI have no answer to a fairly basic governance question: which agent did what, under whose authority, and can you show me. The answer is currently “check the logs of whichever vendor we happened to buy”. That holds fine while agents are drafting emails. It stops holding the moment an agent touches money, customer records, or production.
Buzz’s answer is portable cryptographic identity per participant and a signed event log. Whether Buzz is the product that wins does not much matter. The question it is answering is one every serious tool is going to have to answer, and most of them have not started.
The second signal is structural. Buzz is an open protocol, self-hostable, with no per-seat billing, launched by a company large enough that nobody can dismiss it as a hobby project. That is a direct challenge to how workplace software has been priced for fifteen years. Even if Buzz stays niche, its existence changes what a team can reasonably ask for.
What to do this week
Do not migrate. Install it.
Pull down the desktop client, spin up a workspace that does not matter, and put one Claude Code agent in a channel with one real task. Give it half an hour. What you are looking for is not whether the app is good yet, it is whether working alongside an agent that has its own identity and its own signed history feels different to working with a bot you installed. In our experience it does, and once you have felt it, the way agents are bolted onto your current tools starts to look temporary.
Then write down what you would need before you would trust it with real work. Approval gates. A stable release. Whatever your own list holds. Keep it somewhere you will see it again, because the gaps are closing faster than you would expect: mobile went from missing to shipped on two platforms in eight days. When Buzz or something like it ticks your list, you will be six months ahead of the businesses that only started thinking about it then.
