Buzz Puts AI Agents in a Shared Team Workspace

There's a coordination problem baked into how most people use AI agents right now. You have Claude Code doing one thing, Cursor doing another, maybe Codex sitting in a third tab. Each one is capable. None of them knows what the others are doing. You, the human, become the integration layer—copying outputs, pasting contexts, manually handing work from one tool to the next. It's the kind of irony that should bother anyone who got into AI to stop doing that sort of thing.

Buzz is Block's answer to this problem. The open-source platform—backed by Jack Dorsey and framed by its creators as a "native workspace for human plus agent teams"—takes the Slack mental model and extends it to non-human participants. Channels, threads, direct messages, inboxes: all of it, but populated by both people and agents sitting side by side, treating agents as first-class team members rather than tools you invoke and dismiss.

A recent tutorial from AI Foundations walks through the full setup, from spinning up a server to watching two agents hand work off to each other without human intervention. It's worth taking seriously—not as a sales pitch, but as a window into what this category of tooling actually looks like when someone uses it in the real world.

The architecture matters before the demo does

The tutorial creator makes a point early that's easy to skip past if you're impatient for the cool stuff: Buzz is not a new AI model. It's not a new agent runtime. It's an environment layer—a place where the runtimes you're already running (Claude Code, Codex, Cursor, Goose) can coexist and communicate.

That distinction changes how you evaluate it. You're not comparing Buzz to GPT-4o or Claude Opus on capability benchmarks. You're asking a different question: does a shared workspace make a collection of agents meaningfully more useful than those same agents operating in isolation? The demo suggests yes, with caveats.

The practical setup involves a choice with real trade-offs. You can run Buzz locally, on your laptop, for free. Or you can pay roughly $9 a month for a virtual private server (VPS) that keeps the environment running continuously. The tutorial walks through the VPS path, and the reasoning is straightforward: a laptop-based setup goes offline whenever the lid closes. "Your agents can stay online working for you 24/7," the tutorial notes, "and this environment is always on."

That framing—always-on agents—is the core pitch. If you're paying $200 a month for Claude Code, the argument is that idle time is waste, and Buzz on a VPS turns your AI subscriptions into something closer to staff that doesn't clock out.

What the demo actually shows

The tutorial builds a two-agent content team live: Nova, configured as a research/ideation agent with access to VidIQ for keyword research, and Atlas, configured as a scripting agent running Claude Sonnet. They're grouped into a "content squad" and dropped into a dedicated channel.

The human gives a single instruction—research video ideas around running a one-person business using Claude Code—tags the content squad, and steps back. What follows is the part that's genuinely interesting to watch, not because it's magic, but because the failure mode is visible too.

Atlas initially starts working before Nova has finished researching, which the tutorial creator flags as a problem: "I don't think Atlas, the scripting agent, should go to work until Nova actually pulls in some ideas." The agents work it out, but the moment points at something real—agent coordination without tight role sequencing can produce redundant or premature work. The tutorial's response is to note you might need better instructions. That's correct, but it's also worth naming: the quality of an agent team is still upstream of the instructions you write for them.

When the sequencing does work, the result is legitimately compelling. Nova surfaces a title that scores a 97 on VidIQ's keyword metric, posts the brief to the channel, and tags Atlas. Atlas picks it up, drafts a full script grounded in the channel's actual performance data, and tags the human for review. The whole loop runs without the human touching it. As the tutorial puts it: "I don't have a content team. I don't have a research assistant. I don't have an ops manager. What I have is Claude Code."

The audit trail matters here too. Every agent action is signed and logged—you can see exactly which agent did what, in what order. For anyone managing agentic workflows at scale, that accountability layer isn't cosmetic. It's how you debug, and eventually, it's how you trust.

The costs that don't appear in the pricing table

The $9/month VPS cost is accurate. But that number is incomplete by design—it's the floor, not the ceiling.

If you're running Claude Opus for your ideation agent and Sonnet for scripting, you're drawing from your Claude Code subscription tokens every time those agents are active. An always-on multi-agent setup could meaningfully accelerate token consumption, depending on how aggressively you task your agents and how often they're responding to each other versus sitting idle. The tutorial doesn't quantify this, which is understandable for a beginner walkthrough but worth flagging for anyone doing the math on actual costs. Notably, token costs can get brutal at scale—especially when agents are actively communicating with each other throughout the day.

There's also the question of what it means to have your Claude Code tools, MCP servers, and connected integrations (Gmail, Google Calendar, VidIQ, Slack) accessible to agents running in a shared workspace on a third-party VPS. The tutorial briefly notes that agents inherit all your connected tools automatically, which is convenient and also a meaningful permissions surface. If an agent is misconfigured or given unclear instructions, it has access to the same tools a well-configured one does. That's not unique to Buzz—it's true of any agentic setup—but the always-on, multi-agent context raises the stakes compared to a single session you're watching in real time.

What it actually replaces

The tutorial ends on a framing I find worth examining. The claim is that Buzz replaces a content team—a research assistant, a scriptwriter, an ops manager. That's probably true for a solo creator at a certain scale. It's less obviously true for a team that already exists and has institutional knowledge, client relationships, and judgment that doesn't reduce to a system prompt.

The more accurate frame might be: Buzz replaces the human-as-integration-layer for people who are already using multiple AI tools. If you're already paying for Claude Code and Cursor and you're already doing the manual handoffs between them, a shared workspace with persistent memory, signed audit logs, and agent-to-agent communication is a real upgrade. The upgrade is in coordination, not in capability.

That's worth knowing before you set it up. The agents are only as smart as the models powering them and the instructions you've written. Buzz makes those agents easier to coordinate and keeps them running when you're not at your desk. What it doesn't do is make a poorly-configured agent team into a good one.

The question for anyone considering it isn't really "is this impressive?" The demo makes a reasonable case that it is. The question is whether the overhead of setting up a well-structured, well-instructed agent team—the real work, which the tutorial glosses over—produces enough consistent output to justify the ongoing cost and the expanded permissions surface. That's a math problem every user has to solve for their own situation.

---

Rachel "Rach" Kovacs covers cybersecurity, privacy, and digital safety for Buzzrag.