Block shipped Buzz on July 21st. It looks like Slack, it works like Slack, and the difference is that your AI agents are members of the workspace, not integrations you connect. They have their own profiles, their own identities, their own channel memberships. You @ them the same way you @ a team mate.
I spent a good amount of hours in it by now.
The Good
Agents are members, not plugins. This sounds like a marketing distinction until you use it. An agent sits in a channel, reads what was said before it was tagged, and answers with that context already loaded. If you hate re-explaining, you are going to love this.
The harness underneath is swappable. Every agent has a dropdown for what actually runs it: Claude Code, Codex, and others. Switching does not reset the conversation.
It works on my phone with real threads. I have been talking to agents in Telegram for months, which works right up until three conversations collide in one flat message list. Buzz has proper threading and a mobile app, so I can kick off work from a train and read the result in its own thread.
What did not work
The automations were flaky. These are the trigger-and-action rules (when a message matching X arrives, do Y). My first several attempts to set one up did not fire. They work now, after an update, but this all still seems a bit janky.
Sharing an agent with a teammate failed. The pitch is that a colleague can tag an agent running on my machine and my subscription. I tried it with a coworker and could not get it working.
It is slower than the CLI. Everything routes through your relay and back, so an agent in Buzz responds noticeably slower than the same model running in a terminal on the same machine. For a quick fix, I still open the terminal.
It is the wrong tool for serious engineering work. Small apps, prototypes, internal tools: fine. A real codebase with real complexity: not yet.
Where that leaves it
This is a developer preview and it behaves like one.
What I keep coming back to is that the idea is correct. Your team is going to include agents, those agents need the same context your humans have, and the tool that holds that context should not be one that makes it impossible to leave. Buzz gets that part right before it gets the reliability right, which is the better order to get things in. I wouldn’t say you need to jump on this now, but I do hope that they improve Buzz and turn it into what it could become.