Buzz Puts AI Agents on the Org Chart

On July 21, Block released a workspace application called Buzz, and the positioning was not subtle. Jack Dorsey's company had built a tool intended to make two of the most entrenched products in corporate software, Slack and GitHub, less necessary to the people who use them every day.

That is an ambitious claim. Salesforce paid nearly $28 billion for Slack. Microsoft owns GitHub. Both sit at the center of how modern companies talk to each other and ship work. Displacing either is not a weekend project.

But the argument behind Buzz is worth understanding even for readers who will never install it, because it is not really an argument about chat software. It is an argument about who, or what, is allowed to be a member of a team, and how a business proves after the fact who did what.

What Block actually shipped

Buzz is a group workspace that puts messaging, code hosting, project workflows, and AI agents in the same place. Desktop clients are available for macOS, Windows, and Linux. The source code is published on GitHub under the Apache 2.0 license, which means anyone can read it, run it, or fork it. Teams can host it themselves or use Block's hosted version at buzz.xyz, for which no price has been announced.

The unusual part is underneath. Buzz is built on Nostr, an open protocol where identity is a cryptographic key pair rather than an account on somebody else's server. Every message, code change, and automated action is a signed event written to a shared log. The chat channel and the code repository are not separate products talking through an integration. They are two views of the same record.

The other unusual part is what happens to AI agents in that design. In Slack or Teams, an AI assistant is a bot with an API token, or it borrows a person's credentials. In Buzz, an agent gets its own key, its own membership, and its own permissions. It joins channels, reads history, reviews code, proposes changes, and answers questions the way a colleague would. Its actions are signed, so they can be traced back to the agent and to the person who authorized it.

Buzz does not care which model sits behind the agent. Block says it works with Anthropic's Claude Code, OpenAI's Codex, and goose, Block's own open-source agent framework, among others.

The bet Dorsey is making

Dorsey has spent most of a decade backing open protocols over proprietary platforms, and Buzz is the most direct version of that bet so far. Coverage of the launch framed it as a challenge to Salesforce and Microsoft, and Block has said it intends to use Buzz internally to reduce its own dependence on Slack and GitHub.

The reasoning goes something like this. Individual output has risen quickly now that people can hand real work to AI agents. Team coordination has not kept up, because the context an agent needs is scattered across a chat tool, a code host, a project tracker, and a document library, each with its own permissions and its own idea of who is allowed to see what. Tools that treat AI as an add-on inherit that fragmentation. A tool designed around agents from the beginning does not have to.

Whether that is worth changing platforms over is a separate question. But the diagnosis is not wrong, and it is not confined to engineering.

The part marketers should notice

Marketing departments are already running agents. They draft copy, summarize customer calls, sort inbound messages, adjust budgets, answer first-line questions, and increasingly take actions that touch revenue.

Very few of those departments can answer a plain question about them: who did this, and who said it was allowed?

When an agent writes the subject line that goes to two hundred thousand people, the record usually shows a person's name, because the agent ran under a person's login. When an agent updates a customer record, changes an offer, or replies to a prospect, the trail often ends at a shared account or an integration key. If something goes wrong, a claim that should not have been made, a discount that should not have been offered, a message sent to a list that should have been suppressed, the company is left reconstructing the story from memory and from logs that were never designed to answer the question.

This is the idea in Buzz that travels furthest beyond engineering. Giving an agent its own identity, tied to the person who provisioned it, turns "the AI did it" into something a business can actually examine. It makes an agent's access revocable without revoking a person's. It makes the audit trail honest.

Regulated marketers already understand why that matters. Anyone who has sat through a compliance review of an email program, a claims substantiation exercise, or a data-processing audit knows the question is never only what was sent. It is who approved it, on what basis, and whether you can show your work.

The objections are real

Buzz launched to a genuinely split reception, and the skeptics raised points worth keeping.

The most substantive concern involves the same shared context that makes the product appealing. If an agent sits in many channels and can read all of them, it can also carry something from a channel one person is cleared for into a conversation with somebody who is not. That is not a flaw unique to Buzz. It is an unsolved problem in every multi-agent workspace. Giving agents identities makes their actions traceable. It does not by itself decide what an agent should be willing to repeat.

Developers also questioned whether the decentralized protocol earns its place, or whether a conventional shared service would do the same job with less novelty attached. Others found the premise of AI agents as colleagues unappealing on its own terms and dismissed the demonstrations as chat rooms full of bots.

And the product is early, which Block says openly. Mobile clients and parts of the enterprise feature set are still in progress, the code hosting is young, and the hosted version has no published price. Very few marketing organizations should be evaluating this as a Slack replacement today.

What to do with it

The useful response is not to pilot Buzz. It is to take the question the product is built around and point it at your own stack.

Ask which agents are running in your marketing operation right now, and under whose credentials. Ask whether an agent's actions in your CRM, your email platform, and your ad accounts are distinguishable from a person's. Ask who can switch a given agent off, and how quickly. Ask what the record would show if a regulator, a client, or a customer asked you to explain one specific message six months from now.

Most teams will find the answers thinner than they assumed. That gap is not a reason to panic, and it is not a reason to stop using agents. It is a reason to start treating them the way you already treat contractors and junior staff, with defined access, a named owner, and a record of what they did.

Buzz may or may not loosen Slack's hold on anybody. The idea underneath it, that an AI agent doing real work should have a name, a manager, and a paper trail, is going to outlast the argument about which chat application wins.