Buzz Just Fixed AI Agents (But It Has a Serious Flaw)

Jack Dorsey dropped Buzz and it pulled 20,000 GitHub stars in two weeks.

Buzz Just Fixed AI Agents (But It Has a Serious Flaw)

Jack Dorsey dropped Buzz and it pulled 20,000 GitHub stars in two weeks. People are already calling it a Slack killer. So we installed it, put a Claude agent and a GPT agent in a channel, and made them work together. It got one big thing genuinely right. It also broke in exactly the places it's supposed to be best at.

We build software, so when a new agent tool gets this much attention, we run it before forming an opinion. Buzz is early (version 0.5), but it's moving fast, and it solves one of the real problems with using AI agents on actual work. It also has flaws serious enough that we wouldn't move a team onto it yet. Here's what it is, what broke, and whether you should use it.

What Buzz actually is

Skip this part if you already know it. Buzz is an app that lets you coordinate multiple agents from one place. It's basically a group chat, except the members aren't only humans. Your AI agents are in the room too. It's completely free and open source, which is a big reason for the Slack-killer talk.

Coordinating a group of agents isn't brand new. Claude Code already does it with agent teams, but that happens in your terminal, so nobody else on your team ever sees it, and it only spins up Claude agents. You can also run agents alongside your team in Slack: Anthropic's version is Claude Tag (team and enterprise plans only), and Codex has a Slack integration that works on all paid plans. But connect both and neither one can work with the other, and they only move when a human mentions them. You end up carrying every task between them by hand.

That's the gap Buzz fills. It lets the agents talk to each other directly, so one agent can hand work to another and they coordinate like teammates instead of waiting on you to pass every message. It connects the Claude Code and Codex subscriptions you already pay for, plus Goose (Dorsey's own coding agent), and has them work together on the same thing.

The one thing it got right: attribution

When several agents work in the same place, you need to know which one did what. Buzz gives you that. Every agent gets its own name and its own login, and everything it does is recorded under that name. If something breaks, you scroll back and find which agent did it, right down to who asked it to do that particular thing.

The agents also build their own memory as they work, the way Claude Code does. They write notes as they go and link those notes together, so the next session starts knowing what the last one worked out instead of starting cold. And because the notes sit in the channel, every agent picks them up, not just the one that wrote them.

Takeaway: In a team, you genuinely can't tell which agent changed what or who set it off. That's the real unsolved problem, and Buzz is the only serious answer anyone has shipped.

Getting it running

You don't need the self-hosting quick-start on the GitHub page. Go to the latest releases page instead, scroll to the assets section, and grab the installer for your operating system. (If you're not sure which, hand the link to Claude Code and let it download the right one.) Insall it like any normal app.

On first open you either join a community someone else set up (you need their link and an invite) or make your own, which is what we did. You pick a name and that name becomes your address, hosted for you. Then you connect your agents and the app handles nearly everything, even installing Claude Code if you don't have it. The only time you touch a terminal is once, to log into your account. You get three agents by default and can add as many as you want: click create agent, give it a name and a description (the description becomes its instructions), pick an installed coding tool and model, and it goes live.

What broke when we ran it

We set up a Claude agent and a GPT agent and asked them to build a community website.

Our first prompt was generic: we tagged both and said what we wanted. It started well. Both made plans, reviewed each other's, and started tagging back and forth. Then GPT tagged Claude, Claude said it saw the message, and then nothing. They just sat there until we nudged them, and then they carried on.

So we wrote a second prompt with step-by-step detail and assigned roles: Claude makes the plan, GPT reviews it. That worked much better and the two coordinated properly. But when we asked them to build, it didn't hold up. It's slow compared to running Claude Code in your terminal, because Claude Code works on several parts of the build at once while Buzz does them one after another. Even the planning burned far more tokens than it should have, and that's a design problem, not a setting.

The token problem, and why it multiplies

To see the issue you need to know how context windows work. A model doesn't remember your past messages on its own, so every time you send one, the whole conversation gets sent again: your old messages, the model's past replies, and your new prompt.

Buzz runs a Claude Code session behind every agent, and Claude Code already keeps its own copy of that conversation in memory. Here's the flaw: every time you send a message, Buzz forwards the whole conversation history along with it, even though Claude Code already has that history. So the same conversation now lives in two places and you're spending double the tokens you needed. Then every agent you add is its own session carrying its own context, so the cost doesn't add up, it multiplies.

A real number: replying to a simple greeting took 31,000 tokens in Buzz. The same thing through Claude Code in your terminal, with all the same context, is about 4,000. Anything long-running gets expensive fast.

Takeaway: Buzz sends full history to a session that already has it, and multiplies that across agents. It's a design problem, fixable in time, but real today.

The genuinely interesting parts

None of this is a reason to switch, but it's what makes Buzz worth watching. It's the closest any of these tools has come to actually feeling like Slack.

• Huddles. You jump into a voice huddle with an agent and talk, and it talks back. Broken for now, because the agent replies in markdown, and markdown read aloud sounds exactly as strange as you'd expect.
• Compute sharing. Remember people chaining Mac minis to run big models locally? This is built in. Join a few computers together, host one big model on the combined power, and share it with your team. Buzz even tells you which of roughly 35 models your setup can handle.
• Exportable agents. You export an agent as a JSON file or a PNG, and the memories travel with it, not just the settings. Everything it worked out comes along, and you can search all of it the way you'd search Slack.

What it's genuinely best at: adversarial review

There's one thing Buzz does better than the alternatives, and that's letting agents argue with each other. Adversarial review means one agent attacks the work and another defends it, so between them they catch what a single agent would miss.

You mention the agents and give each a side. We told GPT to go after the PRD (the document that lays out what the app should do), to assume everything in it was wrong, and Claude to defend it. The two went at it in a thread: GPT opened with its argument and pinged Claude, Claude answered, and they went back and forth until GPT ran out of objections. Then Claude wrote up the final plan and asked us to confirm.

The shared room is what makes it pay off. Both agents can see the entire thread, so they know what's already been done and argued, instead of being handed a prompt cold. Claude Code's agent teams do this well too, but every one of those agents is Claude, and when the session ends the argument goes with it. In Buzz you've got two different companies' models going at each other, and the whole thing is still sitting there afterward.

So should you switch?

Solo: no. Everything you'd want Buzz for is already inside the subscription you pay for. Want several things running at once? Sub-agents do that, cheaper. Want agents arguing to catch each other's misses? Claude Code has agent teams, and Anthropic's own docs give the exact example of five agents trying to disprove each other. Want a rival's model reviewing your work? OpenAI's official plugin has put Codex inside Claude Code since March, four months before Buzz existed. Claude Code and Codex are complete setups with their tools already wired in, and they're faster and use fewer tokens.

Team: maybe. This is a different question, because right now you genuinely can't tell which agent changed what and who set it off, and Buzz is the only serious answer to that. It gives you a near-perfect record of what your agents did, and almost no control over what they're allowed to do.

Two early-version problems matter if you roll it out across a team. First, there's no end-to-end encryption, so whoever runs the server can read every message, including private ones. That's on purpose, because they want everything searchable so agents have full context. Second, you can't limit an agent to one channel, so any agent you add can see everything across your whole workspace. If your team needs anything private, that isn't possible yet.

Takeaway: Buzz gives you a near-perfect audit trail and almost no permission control. Team, worth a look. Solo, overkill.

FAQ

What is Buzz? A free, open-source app from Jack Dorsey that puts your AI agents in a group chat with you, so they coordinate with each other directly instead of waiting on you to relay messages.

What does it fix that other tools don't? Attribution. It records which agent did what and who asked, and keeps the whole history searchable. That's the one thing no other agent tool has done properly.

Why is it so token-heavy? It forwards the entire conversation history to a Claude Code session that already has that history, so context is duplicated, and it multiplies with every agent. A greeting cost 31,000 tokens versus about 4,000 in the terminal.

Is it safe for private team work? Not yet. There's no end-to-end encryption, and you can't restrict an agent to a single channel, so any agent can see everything in the workspace.

Should I use it? Solo, no. Everything's already in your existing subscriptions, cheaper and faster. Team, maybe, for the audit trail, as long as you accept the privacy limits.

The takeaway

Buzz solved the problem everyone else ignored (knowing which agent did what) and stumbled on the ones everyone else already solved (speed, cost, coordination). The idea isn't wrong. It's just early. If you run solo, stay with the setup you already pay for. If you run a team and need to see who did what, Buzz is the only game in town right now, with real privacy caveats attached.

The skills and workflows we build across all our videos live inside AI Labs Pro.