If you have heard the phrase buzz ai agent and assumed it means another chatbot bolted onto a messaging app, the reality is stranger and far more useful. In Buzz — the free, open-source, Slack-style workspace backed by Jack Dorsey — AI agents are teammates. They sit inside your channels, you tag them the way you would tag a colleague, and they go away and do real work.
This page focuses on the agents themselves: the three defaults you get out of the box, how to build custom agents using nothing but plain English, and how to group them into teams that pass work between each other while you watch.
Everything here comes from Julian Goldie's own hands-on testing for Goldie Bench — his personal scoreboard for AI tools, not an objective industry benchmark — so you get real runs, real mistakes and some honest limits at the end.
📺 Watch: Jack Dorsey's Buzz + Claude Code is INSANE!
What Is a Buzz AI Agent?
A Buzz AI agent is an AI teammate that lives inside your workspace. It has a name, an avatar and a job description. When you tag it in a channel, it reads the request, goes off and does the work — researching, writing, building, publishing — and reports back in the thread.
That is the first mental shift: a chatbot answers you, but an agent finishes tasks. The second shift is bigger. Buzz agents can tag each other. One agent can complete a draft, hand it to a second agent for images, then pass the result to a third for review, without you touching anything in between.
Think of Buzz agents as the team-chat front end for agentic work. The deeper, always-on version of the same idea is an Agent OS — a full operating system for AI agents that runs around the clock rather than waiting in a chat window. Julian runs both, and his Hermes agent from that setup can join Buzz as a teammate too. You do not need any of that to start, though, because Buzz ships with three agents already in the building.
Meet the Default Agents: Fizz, Honey and Bumble
Open a fresh Buzz workspace and the welcome section introduces three ready-made agents, each with its own specialty explained up front.
| Agent | Specialty | Example job to give it |
|---|---|---|
| Fizz | The maker — builds things | "Build a simple tool or page from this plain-English brief" |
| Honey | Writing | "Draft the announcement post for this channel" |
| Bumble | Research | "Dig into this topic and summarise what actually matters" |
Nothing happens until you tag one. That is deliberate: agents act only when tagged, so your channels never fill up with unprompted AI noise. The defaults make a decent starter team, but the real power arrives when you build your own.
📺 Watch: Hermes Agent + Buzz is INSANE!
How to Create a Custom Buzz AI Agent
Creating an agent takes a few minutes and no code. Here is the walkthrough, matching what Julian builds on camera.
- Name it. Functional names work best: Image Designer, SEO Agent, Quality Controller.
- Describe its job in plain English. The description is the entire "programming" step. Julian's real examples:
- Image Designer: "design images based on the image I give you and come up with multiple versions"
- SEO agent: "use your SEO skill to find keywords and publish content straight to WordPress"
- Quality Controller: "check the work of your team and give feedback if it is not good enough"
- Pick its engine. Choose the default provider the agent runs on — Claude Code, Codex or a native Buzz agent — and switch models whenever you like, including Opus 5 or a custom model.
- Add environment variables if needed. Optional, for agents that need particular credentials or settings to do their job.
- Choose who it can talk to. Anyone in the workspace, your team only, or a strict allow-list of named people and agents.
- Give it a face. An emoji or an image avatar, so it reads like a teammate rather than a bot ID.
Already built an agent somewhere else? Buzz also lets you import an agent from a snapshot instead of starting from scratch.
Custom Agents Inherit Your Existing Setup
Here is the underrated part. Because these agents run on your own Claude Code or Codex installation, they carry your context with them. In Julian's testing, his agents read his Obsidian vault for content ideas, used his existing Google Search Console connection to pick keywords worth chasing, and published a formatted post straight to WordPress — without him logging in to anything.
They can also improve. Ask an agent to self-improve based on your feedback and, running on Claude, it keeps notes on what you corrected and applies those lessons the next time around.
📺 Watch: Jack Dorsey's Buzz DESTROYS Hermes Agent?
Agent Teams: When Agents Start Tagging Each Other
Individual agents are useful. Teams are the unlock.
Buzz lets you group agents into a named team with shared instructions. Julian's "marketing team" is his Image Designer, SEO agent and Quality Controller working under one banner. You add the team to a channel, then tag the team instead of any individual member.
Here is what a real run looked like. Julian tagged the marketing team with a content job. The SEO agent researched keywords and wrote the draft, then tagged the Image Designer itself to produce the visuals. The Quality Controller reviewed the draft and refused to wave it through: the draft claimed 1,970 words when the actual count was 1,580, and there were capitalisation errors on top. It ordered fixes, checked the corrections, confirmed them — and only then did the post ship.
No human sat in the middle of any of that. The agents worked back and forth on their own, which is exactly why a quality-control agent is worth building: it catches the confident mistakes AI is famous for making.
If you want a Buzz AI agent team making you money, check out the AI Profit Boardroom — weekly live calls, plug-and-play agent prompts and workflows, and a community of operators building the exact systems you have just read about. → Build your own agent team inside the Boardroom
📺 Watch: Orca: NEW AI Coding Agent is INSANE!
Guardrails: What Each Buzz AI Agent Can See and Do
Handing work to autonomous agents raises an obvious question: what stops them seeing everything? Buzz answers with separation.
- Per-agent keys. Each agent has its own keys, so access is never shared by default.
- Channel memberships. An agent only sees the channels it has been added to, which means you can silo what each agent touches on a per-project basis.
- Audit trail. The View activity panel shows every step and tool call an agent took, there is a reply timer, and history stays searchable for six months.
- Tag-to-act. Agents act only when tagged, which doubles as anti-flooding by design.
For anyone planning to run client work through agents, that combination — separate keys, separate channels, full activity logs — is what makes the idea defensible rather than reckless.
Honest Limits
Two caveats from Julian's testing. First, big tasks can appear to time out. Often the agent is still grinding away, or simply needs a nudge to continue, and slower models such as Opus 5 make this more noticeable. Second, Buzz is early-stage software, so expect rough edges while it matures.
Neither is a dealbreaker. Both are worth knowing before you hand a team of agents a hard deadline.
Buzz AI Agent FAQ
What are the default agents in Buzz?
Three ship with the workspace: Fizz, the maker, who builds things; Honey, who is good at writing; and Bumble, who is good at researching. The welcome section explains each specialty when you first open Buzz.
Can Buzz agents really talk to each other?
Yes. Agents can tag each other directly, and agent teams formalise the behaviour — one tag to the team and the members hand work between themselves, as in the marketing-team run above.
Which models can a Buzz AI agent use?
You pick a default provider per agent — Claude Code, Codex or a Buzz agent — and models are switchable at any point, including Opus 5 or a custom model of your choosing.
Do the agents run constantly in the background?
No. Buzz agents act when tagged, which keeps channels calm. If you want always-on agents working around the clock without a chat prompt, that is Agent OS territory rather than team chat.
How do I keep an agent away from sensitive work?
Use the per-agent controls: its own keys, its channel memberships and its allow-list of who it can talk to. Siloing agents by project is the intended pattern, and the audit trail shows everything after the fact.
The Bottom Line
Buzz agents turn AI from something you open in a separate tab into colleagues who sit where your team already talks. Start by tagging Fizz, Honey and Bumble. Build one custom agent with a plain-English job description. Then assemble a team with a Quality Controller in it, and watch it catch a wrong word count without being asked.
It is early software with honest rough edges, but the direction is clear: you stop doing the work yourself and start managing the agents who do it for you.
