Build a Project Room Where Humans and AI Agents Work Together (Buzz Tutorial)

Go to buzz.xyz (https://buzz.xyz), click Get the App, and install Buzz. Open the app and choose the coding harness you already use, such as Codex or Claude Code.

Buzz uses that account to power the agents in your workspace, so make sure its command-line tool is installed and authenticated before continuing. During setup, choose a model or accept the recommended defaults. You can adjust agent models and other defaults later from the app's settings.

*Choose the coding harness you already use, then select a default model for new agents.*

If you are not using the same Mac setup shown in the video, check the official Buzz repository (https://github.com/block/buzz) for the latest packaged releases and platform instructions.

Buzz begins with an onboarding conversation. Tell the onboarding agents about your role, the work you manage, and the main projects you want help with. Then ask them to organize those projects into separate chat channels.

A useful instruction looks like:

I manage several projects that involve research, content, and simple web builds. Create one channel for each active project, and add me to every channel you create.

Treat a channel like a project room rather than a general-purpose AI chat. Keeping one outcome per channel makes its messages, files, and agent activity easier to follow. If an agent creates a channel without adding you, ask it to add you before the work begins.

*Each channel keeps the people, agents, messages, and files for one project in the same room.*

Channels can also be private to their invited members. That makes it possible to separate internal work, client work, or individual projects inside the same Buzz workspace.

Open Agents and review the agents created during onboarding. For this workflow, keep the team simple:

• A lead agent that receives your instructions and coordinates the project
• A research agent that gathers and summarizes the case studies
• A maker or coding agent that can turn the research into a webpage

You can ask the lead agent to create a dedicated research agent, or create one yourself from the Agents area. Choose a recognizable name and icon, review its instructions and connected tools, and add it to the project channel. An agent cannot participate in a channel it has not joined.

Review each agent's permissions before giving it work. Buzz keeps agents in a sandbox by default during the demonstrated workflow, and access outside the workspace should require your approval. Limit who can instruct an agent when a project contains sensitive context or when you want tighter control over token usage.

For your first project, avoid an elaborate roster. A lead, a researcher, and one maker are enough to learn how delegation works. You can turn a successful workflow into a more specialized reusable team later.

In the project channel, address the lead agent and give it an outcome with two independent parts. Use this prompt from the demonstration:

Have the Research Agent find five useful AI productivity case studies. While it researches, draft a webpage that presents each one consistently.

This tells the lead agent what should be delegated and what can happen in parallel. The research agent can collect the source material while the lead or maker begins the page structure.

*Buzz shows both agents working beneath the project prompt, making the parallel handoff visible.*

Make sure the agent mention resolves before you send the message. Type the @ symbol, select the agent from the menu, and confirm its name appears as a mention. If you type an agent's name without tagging it, the agent may never activate. You can also enable the setting that keeps addressed agents active when you expect a longer exchange.

If you want a dedicated coding agent to own the build, say so explicitly. For example:

Delegate the research to the Research Agent and the webpage build to the coding agent. Keep the page simple and use a familiar stack.

Starting with a simple, familiar stack reduces the chance that multiple agents choose incompatible libraries or create unnecessary setup work.

Follow the agents inside the channel as you would follow human teammates. Reactions show when an agent has seen a message, and the activity view shows what it is doing. Threads keep follow-up questions attached to the relevant update.

Ask for a status update when you need one. If an agent chooses an unsuitable tool, library, or approach, correct it in the channel and restate the constraint. The shared history gives the team the context needed to continue without starting a new chat.

When the work is ready, ask the agent to attach the files or screenshots directly in the channel. Review the case studies, confirm that each one follows the same structure, and test the webpage before treating the project as finished. Visible activity is useful, but it does not replace checking the final output.

Keep agent permissions narrow as the project grows. Approve access to other folders only when the task requires it, and verify important source claims or code changes before sharing the work outside the project room.