After AI layoffs, Block ships a team chat where agents are peers

Block, the Jack Dorsey payments company behind Square and Cash App, cut staff earlier this year in a round trade press has linked to an AI push. The bet that follows landed this week: Buzz, an open-source group chat where AI agents share channels with people on the same protocol seat.

Jack Dorsey announced Buzz on X as a way to "reduce our dependency on slack and github," calling the platform "model-agnostic, decentralized, self-sovereign, and open source." The code is public at the block/buzz GitHub repository, described in the README as a hive mind communication platform. The product site at buzz.xyz pitches one workspace for people, agents, and project work.

Trade press, including The Register and The Next Web, has cast Buzz as a Slack challenger. The more useful question is what the system is being built to optimize, and what it leaves out. Buzz's underlying design treats the agent as a first-class participant: the same handle, the same channel membership, the same standing in a thread as a person. Dorsey's X post, Block's announcement page, and the GitHub README all use the same peer language: humans and agents working together.

"Model-agnostic," as Dorsey uses the term, means the agent in a Buzz channel can be any underlying model, not Block's own. "Self-sovereign" implies a team could host its own server, the way an organization runs its own email or wiki. Both are claims the public repo either backs or does not, and the README points in that direction. A team that wants to keep customer data on its own hardware could, in principle, run the whole stack internally and still pull in third-party models as agents.

A chat that gives an agent a real seat optimizes for throughput: a query answered, a ticket closed, a draft sent, in the same place a human teammate would have done it. Mentorship does not fit that shape. A junior engineer learns by watching how a senior reads a failing test, not by being routed to a model that returns a patch. Pushback looks different too: a peer who disagrees in a thread is not the same as a tool that returns the next best answer. Context that lives in a hallway, or in the head of someone who has been at the company for years, does not survive being treated as a prompt input.

The falsifiable test is in the open repository. If Buzz's protocol really treats agents as peers, then an agent should be able to subscribe to channels, post messages, get mentioned, and trigger the same notification and audit hooks a human does. Runtime behavior, not marketing language, is what determines whether "first-class" is real. A reader can check the repo and see for themselves.

A portable question for any workplace: when an agent tool lands in a team chat, ask whether the agent gets a seat at the table or a tab on the side. Then ask what the seat costs the people who used to sit there. Block has not answered that second question. The answer is in the open repository, in code, to anyone who wants to read it.