Block shipped Buzz on Tuesday — an open-source workspace where AI agents are full team members, not bolted-on bots. Agents get their own cryptographic identity, their own channel memberships, and their own Git permissions. The repo is Apache 2.0 on GitHub. Block’s hosted relay is free in beta. Self-hosting requires Docker, Rust 1.88+, and Node 24+.
The Problem Is Coordination, Not Capability #
Block’s engineering team makes a direct argument in their launch post: “The bottleneck moved from intelligence to coordination.” Models are good enough. The reason AI agents haven’t integrated into team workflows is that Slack lives in one tab, GitHub in another, and agents run as disconnected terminal commands with no shared context. Buzz is Block’s answer to that gap.
Features and bugs each get a short-lived channel. Discussion, patches, CI results, code review, and merge decisions live in the same thread. When the feature ships, you have a complete record — including the failed approaches, the reasoning, and who (or what) made each decision. That last part matters when the “who” is an agent.
Agents Get Cryptographic Identities — Not Borrowed Credentials #
This is the part worth paying attention to. Traditional bots operate on borrowed human tokens. Buzz gives each agent its own Nostr keypair. The agent signs its work. The human owner signs a narrowly scoped authorization linking the two. Block calls this “delegation cryptography.”
The practical difference: if an agent key leaks, you revoke it without touching your own identity. You can audit exactly what an agent did versus what its human owner authorized. And because the identity lives on Nostr — a decentralized protocol — it isn’t tied to Block’s servers. If Buzz shuts down, your signed history still verifies.
“Every company is going to need a place where humans and agents work together. The question is whether that place is proprietary or open.”
Bradley Axen, Head of AI Capabilities at Block
ACP: The Protocol That Connects Agents to the Workspace #
Buzz doesn’t mandate which AI framework you use. Agents connect via the Agent Client Protocol (ACP), an open standard that currently supports Claude Code, OpenAI Codex, and Block’s own goose. ACP is explicitly complementary to MCP — where MCP connects models to external data sources, ACP wires coding agents into editors and workspaces. It re-uses MCP’s JSON representations where it can.
The practical implication: you can swap from Claude Code to Codex to goose without rebuilding permissions or losing channel history. The agent identity and the workspace configuration are independent of which model is running the agent. That’s a meaningful architectural decision in a field where everyone is trying to create lock-in.
Git Designed for Machine-Scale Commits #
Traditional Git forges were built for humans committing code at human speed. Once you have 20 concurrent agent branches pushing patches in parallel, a filesystem-based Git backend becomes a bottleneck. Buzz implements Git on object storage using compare-and-swap atomic pointer updates — a design formally specified in TLA+ and backed by a conformance suite every storage backend must pass.
Block built this because their own internal agent workflows hit the wall on conventional infrastructure. The result handles what Block describes as “human-months of commits and CI runs in an afternoon.” That’s not a feature most teams need today, but it will be.
What’s Missing at Launch #
Version 0.4.22 is not a finished product. Mobile clients are still in development. There’s no desktop app yet. Push notifications aren’t available. Web-of-trust reputation systems — which would let the workspace weight agent outputs by track record — are planned but not shipped. For teams evaluating Buzz today, the hosted beta at buzz.xyz removes the infrastructure burden, but the feature set reflects early-stage software.
Open vs. Proprietary Is the Real Bet #
The enterprise alternative to Buzz is NVIDIA and ServiceNow’s Project Arc — a governed autonomous desktop agent with auditability features, but built on proprietary infrastructure. Buzz is betting the opposite: that an open protocol anyone can implement is structurally safer than a closed platform anyone can discontinue.
Whether Buzz succeeds depends largely on ACP adoption. If the Agent Client Protocol achieves the kind of ecosystem traction MCP has, Buzz becomes foundational infrastructure. If ACP remains niche, Buzz is a well-engineered chat app with an interesting identity model. The code is open. Block’s thesis is on record. Developers will decide.