Is Jack Dorsey's Buzz worth using to manage a team of AI agents?

Buzz for AI agents: a founder's honest first-week test

Buzz, Block's open-source Nostr-based workspace launched July 21, 2026, gives each AI agent its own cryptographic identity so agents can post, review code, and run workflows alongside humans instead of acting as invisible bots. Applying Sir John Whitmore's GROW model from Coaching for Performance, I only grant an agent standing access after a four-step awareness-and-responsibility check, and most default agent setups fail step one.

My Slack was a mess before I even heard of Buzz. Three different agents — a research bot, a Claude Code instance, and a scheduling assistant — all posting under one shared "AI Team" account, with no way to tell which one made which decision when something went wrong. So when Jack Dorsey's Block released Buzz on July 21, 2026 — a free, open-source workspace on GitHub (github.com/block/buzz) that merges team chat, git hosting, and AI agents under the Apache-2.0 license — I gave it a real week, not a demo click-through.

Buzz's actual design choice is more interesting than "Slack with bots." It's built on the Nostr protocol, and every participant — human or agent — holds its own cryptographic keypair, with a second signature linking an agent back to its human owner. That means an agent's history, permissions, and reputation are portable and independently verifiable, not just an API key sitting in a vendor's database. Agents can post in channels, review code, run YAML-defined workflows, and merge into Git branches treated as discussion threads — alongside frameworks like Claude Code, Codex, or Block's own Goose, since Buzz is explicitly model-agnostic. "Every company is going to need a place where humans and agents work together," Bradley Axen, Block's Head of AI Capabilities, said at launch. "The question is whether that place is proprietary or open."

That's a genuinely useful infrastructure decision. But cryptographic identity solves a different problem than the one that actually burned me with my messy Slack setup. Identity tells you which agent did something. It doesn't tell you whether that agent should have been trusted to do it in the first place — and that's a leadership problem, not an engineering one. Sir John Whitmore's Coaching for Performance makes the case that unlocking performance in any team — human or otherwise — comes from raising two things: awareness (does the actor actually see the situation clearly) and responsibility (does the actor genuinely own the outcome, not just execute an instruction). Handing an agent a keypair and admin scope without either of those is just command-and-control with better bookkeeping.

So instead of onboarding my three agents into Buzz the way the docs suggest — connect, grant permissions, go — I ran each one through a version of Whitmore's GROW model before giving it standing access to anything:

1. Goal — what specific outcome is this agent accountable for? Not "help with research" but "draft a competitor pricing summary I can act on without re-verifying every number." Vague goals produce vague accountability, and vague accountability is exactly what made my old setup unreviewable.

2. Reality — what's its actual track record, not its marketing? I pulled two weeks of my Claude Code agent's prior outputs before granting it merge rights in Buzz's Git integration. It had a clean run on scoped tasks and a rough one on anything requiring judgment calls about scope — so scope-judgment tasks stayed manual.

3. Options — standing access or session-scoped? Buzz makes this a real choice, not a binary. My scheduling assistant now runs on session-scoped tokens that expire after each use; only the research agent, with the cleanest track record, got a persistent keypair.

4. Will — who signs off, and where's the audit trail? Every agent's cryptographic identity now logs to a channel I actually read weekly, not just a dashboard I'll forget exists by September.

The failure mode I hit in week one: Buzz's Git integration is, by Block's own admission, "still early," and my research agent posted a workflow update to a shared channel that I didn't notice for a full day because I wasn't watching closely enough — the exact interference Whitmore warns about when responsibility is assigned but awareness isn't actively maintained. A cryptographic identity told me afterward exactly who did it. It did nothing to prevent it. For a solo founder without a compliance team, that gap is tolerable; for anything client-facing, I'd want a human review gate on merges regardless of how clean the agent's history looks.

Compared to my old Slack-plus-Claude-Code setup, Buzz is a real upgrade on traceability and portability — the identity doesn't disappear if I switch model providers. It is not, on its own, a substitute for actually deciding which agents have earned trust. That decision is still mine to make, every week, the same way it always was with human hires.

Sources: Block's official Buzz launch post, Buzz on GitHub, Business Today's coverage of the July 21, 2026 launch, and Sir John Whitmore's Coaching for Performance for the GROW model.