Jack Dorsey's Buzz Isn't a Slack Competitor. It's a Bet on Who Owns Your Team's Memory

Jack Dorsey has launched Buzz, an open-source workspace where people and AI agents work in the same place. His framing: team chat, agents, and project work in one system, model-agnostic, decentralized, self-sovereign, built to reduce a team's dependence on Slack and GitHub.

Read that as a product announcement and it is the hundredth attempt to unseat Slack. Those attempts have a bad record, and nothing in a launch post changes it.

Read it as an architecture claim and it asks a question that is genuinely live right now: when agents do a growing share of the work, who holds the record of that work, and can you take it with you?

The bundle is the claim

The easy way to dismiss Buzz is to call chat plus agents plus project tracking a feature bundle. Every tool eventually grows the neighbouring tool's features. Slack added workflows. GitHub added discussions. Nothing new.

But the bundling is the argument, and it follows from how agents actually fail.

An agent is bounded by the context it can reach. On most teams, the context of a single decision is split across systems that were never designed to be read as one document. The reasoning sits in a pull request review. The decision that overruled it sits in a thread. The current state sits in an issue tracker, and the reason the state is wrong sits in a call nobody transcribed. Humans reconcile that gap constantly, cheaply, and mostly unconsciously.

An agent cannot. So teams write the glue: connectors, sync jobs, retrieval layers, a permissions model stitched from two vendors' permission models that do not agree on what a "member" is. In my reading of how these integrations go, the glue is where most of the engineering effort lands, and it is glue that has to be maintained rather than a capability that compounds.

Putting conversation, code, and project state behind one addressable surface with one permission model is a real structural answer to that. Whether Buzz executes it is a separate question, and an unanswered one.

Three words, three landlords

Model-agnostic, decentralized, self-sovereign. It is tempting to file those under ideology. Each one names a specific party who would otherwise have leverage over you.

Model-agnostic targets the model vendor. If your workspace ships with one provider's agent wired into the surface, your switching cost stops being an API key and starts being the workflow your team has built around that agent's behaviour. The reason to care is not neutrality as a principle. It is that model quality and model pricing are both moving fast, and you want the ability to act on that.

Decentralized targets the host. Coordination that runs on one company's servers is coordination that inherits that company's uptime, jurisdiction, pricing, and appetite for your account.

Self-sovereign, in the usual sense of the term, means the team holds its own identity and keys rather than renting an account inside a vendor's namespace. It is the difference between exporting your data and already having it.

None of these are new virtues. What has changed is the price of not having them.

Why the price changed

When humans did all the work, the workspace was a communication cost. Bad chat software wasted time. Annoying, survivable.

When agents do part of the work, the workspace becomes the execution environment. It is where the memory lives, where permissions are enforced, and where automation gets metered. Vendors have already started pricing AI usage separately from seats, which tells you where they expect the value to accrue.

That shifts what lock-in means. The old version was migration cost: exporting messages, moving repos, retraining people. The new version is that the accumulated context your agents depend on sits inside someone else's product, in their schema, reachable through their API, at their rate limit. You can still export the messages. What you cannot easily export is the thing that made your agents useful, which is the connected record of how your team actually decided things.

That is the bet Buzz is making. Not that Slack is bad software. That the record is about to be worth more than the chat client.

The case against it is strong

Open and decentralized has lost this exact fight before, more than once. Matrix has offered an open, federated protocol for team chat for years. It is genuinely good technology. Most companies still run Slack, because nobody picks a workspace for its protocol. They pick the one their contractors, clients, and new hires are already in.

Switching cost in team software is social, not technical. The migration is the easy part.

There is a second problem, and it is the one I would push hardest on. Open source is not the same as sovereign in practice. Most teams will not run their own instance. They will sign up for someone's hosted version, at which point they have a vendor, an account, and a dependency again. The right to self-host is a real option only for the small number of teams who will exercise it. For everyone else it functions as a promise about the vendor's future behaviour, which is better than nothing and considerably less than sovereignty.

And a bundle has to win on three fronts simultaneously. Beating Slack at chat, GitHub at code collaboration, and a dedicated tracker at project work is a harder opening position than a single wedge that does one thing better than anyone.

What would actually tell you something

The announcement is positioning. It is not evidence. If you want to know whether this matters, the useful checks are narrow and concrete:

• Is it usable without infrastructure work? If the sovereign path requires running your own node, most of the sovereignty is theoretical, and the hosted path will define what the product really is.
• What can an agent see, and who granted it? An agent with workspace-wide read access is a different security object than a bot in a channel. The permission model is the product here, more than the UI.
• Is the data portable in a format someone else can read? Not "can you export it." Can a second implementation ingest the export and reconstruct the graph of decisions? That is the only version of portability that survives a vendor going away.

If those three answers are good, the Slack comparison stops being the point. If they are vague, this is a chat app with agents in it, and there are already several of those.

The question is worth more than the launch

I have not run Buzz. Nothing here says it works, and the honest read of the history is that the odds are against any single project displacing entrenched team software.

But the question it forces is one most teams have not answered, and are quietly answering by default every week they add another agent to a workflow inside somebody else's product.

Two years out, the thing your agents read from will be the most valuable artifact your team owns. Buzz is one bet on where it should live. The bet you should be able to state is your own.