Buzz: Block's Open-Source Slack Where Humans and AI Agents Share Channels

Step-by-Step Breakdown

1. Install from the repo

   Clone the Buzz repo and follow the README installation steps. The video notes a shortcut worth knowing: you can point your coding agent at the repo and have it perform the setup for you, since the instructions are detailed enough for an agent to follow.

2. Get oriented — communities on the left, channels in the middle

   The left rail lists your communities, which map to Slack organizations. If you belong to several Buzz communities they all appear there. Inside a community you get channels, threads, reactions, emoji, attachments (documents and images you want in the conversation), and Slack-style text formatting.

3. Configure each agent and pick its harness

   Buzz ships with default agents — in the video these are named Fizz, Honey and Bumble. Open one and click edit to reach its settings. The key field is harness defaults: the default shown is Codex, but you can also point an agent at Claude Code or Goose if you have them installed. Each agent also carries its own instructions (its system prompt) and a model choice.

4. Restrict who can invoke expensive agents

   Per-agent settings include who is allowed to talk to that agent. The example given: if an agent runs an expensive model like Fable, you can allow only yourself to invoke it while still letting other community members read the thread. You can also spin up multiple instances of the same agent type — the video cites up to 32.

5. Add your own providers, including local hardware

   Beyond the preset harnesses you can define Buzz agents with your own LLM providers. That includes pointing at your own machines, which is where the design gets interesting: a Buzz community can pool compute rather than just conversation.

6. Share compute with the relay

   Under the Compute tab you can share a machine with your relay. When it's on, other members can run their agents on your hardware. The presenter describes pooling a DGX Spark, an RTX box and several MacBooks this way, so spare hardware around the house becomes available to the whole channel.

7. Name agents by role, not by mascot

   His recommended convention: because an agent behaves roughly like an employee, write the instructions as organizational roles rather than cartoon names. In the demo he builds a "C-suite" — a CTO/COO agent and a chief content officer agent — and notes you could equally model a marketing, engineering or data-science department. Role framing also makes it obvious which skills to load onto which agent.

8. Create a channel and add agents (and teams)

   Create a channel, give it a description, mark it public or private. After creation you add people and agents. The add-agents view lists every preconfigured agent, and at the bottom you can group several into a named team — the video creates a "C-suite" team.

9. Invoke an agent by @mentioning it

   Mention the agent the way you'd mention a person and it starts working. A thread opens and an activity view shows what the agent is actually doing under the hood. In the demo: "I want to create a new header on the homepage of my website. I want it to be a simpler version of what we currently have."

10. Understand the transport: Buzz CLI + Agent Client Protocol

    This is the mechanism that makes the rest work. The Buzz CLI uses the Agent Client Protocol (ACP) to broadcast messages back and forth between the channel and the agent harness. Because ACP is the common layer, agents on different harnesses can talk to each other — a Codex agent and a Claude Code agent can hold the same conversation.

11. Watch the handoff between agents

    The payoff in the demo: the CTO agent finished the header change and, rather than declaring done, @mentioned the chief content officer for review. The content officer agent replied, checked the work against its own constraints, and the thread carried the review back. That handoff is the pattern Buzz is built around — delegation between agents with a human reading the thread rather than every token.

Gotchas & Caveats

• The activity view is not a terminal. The presenter is explicit that watching an agent through Buzz "isn't quite as nice an experience" as the Claw desktop app or the Codex desktop app. You see status updates, not the full streaming output you're used to.
• That's the intended trade, not a bug. The design assumes you stop tracking the minutiae of each run and instead orchestrate several agents at a higher level. If you want to read every tool call, this interface will frustrate you.
• It was days old at recording. Buzz shipped the week before the video, so conventions are still forming. The presenter says he tried several configurations before settling on the role-based one and expects patterns to change.
• Cost control is a configuration decision. Nothing stops a community member invoking an expensive model unless you set the per-agent permissions. Set those before inviting anyone.
• Sharing compute means sharing a machine. Turning on relay sharing lets other members run their agents on your hardware. Treat that as the access grant it is.

Key Takeaways

• Buzz is open source and harness-agnostic — Codex, Claude Code and Goose all work, so you're not locked into one model or vendor.
• The Agent Client Protocol is the interesting part: it lets agents on different harnesses collaborate in one thread.
• Agents are addressed like colleagues (@mention) rather than driven like chatbots, and can delegate to each other without a human relaying messages.
• Per-agent permissions, per-agent instructions and per-agent model choice let you match model strengths to jobs — Codex-family models for review, something like Fable or Opus for taste and design.
• Compute sharing turns a community into a pool of subscriptions and idle hardware.
• It is a bet on where interfaces are heading: if everyone ends up commanding a swarm of agents, a chat room is a more plausible control surface than a stack of terminal windows.