Latest / AI Fire Daily / #90 Robin: Jack Dorsey’s "Buzz" is Slack for AI Agents - How to Make Claude & Codex Work in the Same Room

Transcript

#90 Robin: Jack Dorsey’s "Buzz" is Slack for AI Agents - How to Make Claude & Codex Work in the Same Room · AI Fire Daily

• 0:00
  Imagine having the smartest employees in the
• 0:02
  world beat, but they are all locked in separate
• 0:06
  windowless rooms. They literally cannot speak
• 0:09
  to each other. Right. They cannot share notes
• 0:10
  at all. Exactly. And that is how we currently
• 0:13
  use AI. Welcome to the Deep Dive. Our mission
• 0:17
  today is unpacking Jack Dorsey Buzz. It is a
• 0:20
  really fascinating open source project. Yeah,
• 0:23
  it is a free Slack style workspace. It solves
• 0:26
  this exact isolation problem we just mentioned.
• 0:29
  It turns your scattered AI tools into a cohesive
• 0:32
  team. We are exploring how to put Cloud Code,
• 0:35
  Codex, Cursor, Grok, and Kimi into a single chat
• 0:38
  room. It's quite the lineup. Oh, absolutely.
• 0:41
  We are going to look at the mechanics of multi
• 0:42
  -agent collaboration. We will cover data ownership.
• 0:45
  And we will outline the honest limitations of
• 0:47
  scheduled AI tasks. Let us start with the core
• 0:49
  problem here. The modern AI experience is defined
• 0:53
  by tab fatigue. Yeah, that is a great way to
• 0:55
  phrase it. Context is just scattered everywhere
• 0:57
  for you. It lives in Claude, chat, GPT, and cursor.
• 1:00
  Right, and it hides in Slack channels and Notion
• 1:03
  tabs. It is essentially invisible work. You spend
• 1:05
  10 minutes writing a complex prompt. You send
• 1:08
  it off into the void. And you just hope the model
• 1:11
  understands. You are assigning complex tasks
• 1:13
  to completely isolated tools. But Buzz approaches
• 1:18
  this from a totally different angle. It mimics
• 1:21
  an interface you already understand intuitively.
• 1:23
  Exactly. It looks and feels like a normal Slack
• 1:26
  workspace. You have channels, direct messages,
• 1:29
  and tags. But instead of just humans, you invite
• 1:32
  AI agents. They become your real teammates. They
• 1:35
  sit right alongside your human colleagues. The
• 1:37
  crucial difference is you aren't paying for a
• 1:39
  new orchestration tool. Ah, right. Buzz actually
• 1:43
  uses your existing agent subscriptions. I have
• 1:45
  to make a vulnerable admission here. I still
• 1:47
  wrestle with prompt drift myself. I'm constantly
• 1:50
  copy -pasting answers between Claude and Codex
• 1:53
  just to keep them aligned. Oh, yeah. It completely
• 1:56
  drains your cognitive load. It really does. But
• 1:58
  it's one thing to say they share a room. How
• 2:00
  does an open source app actually hijack your
• 2:02
  existing corporate subscriptions without breaking
• 2:04
  them? Well, it securely hooks into your local
• 2:07
  installations. When you message codecs inside
• 2:09
  Buzz, you are genuinely using your real codecs.
• 2:13
  I see. It maintains your existing skills and
• 2:15
  memory. So it's a shared control room for the
• 2:18
  engines you already own. Perfectly stated. And
• 2:21
  that brings us to the technical bridge. The protocol
• 2:24
  making this work is called ACP. Let me define
• 2:27
  that jargon for you real quick. A bridge letting
• 2:30
  chat apps control AI tools running on your computer.
• 2:33
  Exactly. And because of ACP, you are not locked
• 2:36
  into a proprietary ecosystem. An agent in Buzz
• 2:40
  is just an existing tool, plus a system prompt
• 2:43
  and workspace instructions. It is like a plug
• 2:46
  and play suit of armor. You are just swapping
• 2:49
  the pilot inside the suit. But the suit's overarching
• 2:51
  mission remains exactly the same. Right. You
• 2:54
  could build a custom agent named Harry. Harry
• 2:56
  uses Codex to research AI tools. Six months later,
• 2:59
  you switch Harry's engine to Claude Code. And
• 3:01
  the context stays intact. The history remains
• 3:03
  perfectly preserved, yeah. Okay, so we have the
• 3:06
  bridge to bring them into the room. But how do
• 3:08
  we prevent these models from talking over each
• 3:11
  other in the channel? You use a very specific
• 3:13
  hierarchy. You assign a lead agent to manage
• 3:16
  the conversation flow. You assign a lead agent.
• 3:18
  to synthesize the noise into actual decisions.
• 3:21
  That is the absolute killer feature here. Multi
• 3:24
  -agent chats usually get incredibly messy. Agents
• 3:27
  repeat themselves and over explain simple concepts.
• 3:30
  Yeah, they do. Buzz fixes this with a shared
• 3:33
  thread. It uses the lead agent pattern. You just
• 3:35
  tag multiple agents in one message. You give
• 3:38
  them a single directive. Right. You say, one
• 3:40
  of you take the lead, talk through the plan,
• 3:44
  disagree, then execute. One searches the web.
• 3:48
  Another checks old contracts. They actually debate
• 3:50
  in the open channel. Wow. And then the lead agent
• 3:53
  finalizes the decision. This is essentially a
• 3:55
  high -end restaurant kitchen. Oh, I like that
• 3:57
  analogy. Yeah, the lead agent is the head chef
• 3:59
  expediting the ticket. The cheaper models are
• 4:02
  the sous chefs prepping the ingredients so the
• 4:05
  head chef doesn't waste energy chopping onions.
• 4:07
  That is the smart model strategy in a nutshell.
• 4:10
  You pin strong models like Codex or Clawed to
• 4:13
  hard tasks, things like system architectural
• 4:16
  or legal review, and you use cheaper, faster
• 4:19
  models for summaries and first pass reviews.
• 4:21
  Even with just one single Codex subscription,
• 4:25
  you can use a parallelism trick. setting parallelism
• 4:28
  to 5, 10, or 20. Exactly. It lets Codex handle
• 4:32
  multiple tasks simultaneously. It acts like a
• 4:35
  small, synchronized team. That kitchen analogy
• 4:37
  works great for text. Yeah. But what about the
• 4:40
  friction of constantly typing out these complex,
• 4:43
  multi -agent instructions? You actually don't
• 4:45
  have to type them out at all. Buzz lets you manage
• 4:47
  these workflows through audio. Right. We can
• 4:50
  just speak to them directly through voice huddles.
• 4:52
  It moves Buzz away from a standard chat box.
• 4:54
  It becomes an immersive working environment.
• 4:56
  You can start a voice call in a channel with
• 4:59
  Codex and Claude. I have to push back slightly
• 5:02
  on that. Mm -hmm. Human thought processes are
• 5:04
  incredibly messy out loud. We ramble constantly.
• 5:07
  Yeah, you do. How does an AI parse that efficiently?
• 5:10
  Well, that is the brilliant part of the engineering.
• 5:12
  Buzz transcribes the audio, but the agents only
• 5:16
  reply to the actionable parts. Oh, wow. Yeah.
• 5:18
  They systematically filter out all the human
• 5:20
  rambling. So we were just talking about breaking
• 5:22
  the text barrier. And there's an iOS app that
• 5:25
  really expands this. Yeah, you just scan a QR
• 5:27
  code to securely sync your phone. You can send
• 5:29
  a task while walking to get coffee. Codex. Make
• 5:33
  three landing pages. Claude, please review them.
• 5:36
  You return to your laptop, and the work is moving.
• 5:38
  It completely decouples you from the keyboard.
• 5:41
  Buzz also has a projects feature. It acts exactly
• 5:44
  like a Git repository. Agents create local folder
• 5:47
  copies of an application. Right. It keeps them
• 5:50
  from stepping on each other's toes. Claude works
• 5:52
  on the front end, Codex works on the back end,
• 5:55
  then they push the code to GitHub. With all this
• 5:57
  sensitive code and voice transcripts living in
• 6:00
  one app, we have to ask a critical question.
• 6:03
  Where is all this deeply valuable agent context
• 6:06
  physically stored? Buzz avoids the standard corporate
• 6:10
  cloud model. They use a relay architecture instead.
• 6:13
  We keep the data on our own relay, entirely bypassing
• 6:16
  third -party servers. Data sovereignty is non
• 6:18
  -negotiable here. You can self -host the relay
• 6:21
  or use a company like Block. The crucial point
• 6:24
  is your team controls the shared memory layer.
• 6:26
  Because agent context is immensely valuable intellectual
• 6:30
  property. Project decisions, refined prompts,
• 6:33
  deep research. And since it is open source, anyone
• 6:36
  can fork buzz. You could build a visual canvas
• 6:39
  view for PDFs tomorrow. You could turn it into
• 6:41
  a full co -working system. Yeah. But under settings,
• 6:44
  there is a shared compute feature. Buzz detects
• 6:47
  your computer specs. It lets others in the community
• 6:50
  use your hardware to run local agents. Imagine
• 6:55
  an agent economy. People sharing hardware. Agents
• 6:59
  actually paying other agents for compute power
• 7:01
  to solve a problem. That is a wild frontier.
• 7:04
  It really is a massive frontier for decentralized
• 7:07
  computing. But bringing it back down to earth
• 7:09
  for a moment, how can a listener start using
• 7:11
  this today without getting overwhelmed? The onboarding
• 7:14
  is very approachable. You just download it for
• 7:16
  free. You let it auto -detect your installed
• 7:17
  agents, install the ACP connector, and pick a
• 7:21
  default lead agent. Start small, install Buzz,
• 7:24
  connect one lead agent, and build from there.
• 7:25
  That is the best way. Pick the model with the
• 7:28
  most useful context for you. The source is also
• 7:30
  mentioned using OpenRouter. You paste an API
• 7:33
  key to access a massive list of LLMs in one place.
• 7:36
  Yeah, it lets you mix expensive and cheap models
• 7:38
  easily. And what about the management agent?
• 7:40
  Oh, that is a narrow PODEX agent with one job.
• 7:44
  It checks your email, Slack, and Notion at fixed
• 7:47
  times, say 9 a .m. and 12 p .m. It delivers a
• 7:52
  prioritized action list. That sounds incredible.
• 7:55
  But the sources give a reality check on automation
• 7:58
  here. Yeah, scheduled workflows are not fully
• 8:00
  reliable yet. The problem isn't capability. It
• 8:03
  is trust. It really is. An agent might say, it's
• 8:06
  9 a .m., checking email, but failed to execute.
• 8:09
  It's refreshing to hear that we shouldn't just
• 8:11
  hand over the keys to the castle yet. Trust is
• 8:13
  the bottleneck, not technology. Exactly. You
• 8:16
  can create a checker agent to verify completion,
• 8:18
  but that adds mental load. What is the ultimate
• 8:21
  takeaway regarding automated scheduling? Stick
• 8:23
  to supervised workflows. Keep humans in the loop
• 8:25
  for now. Trust supervised peer review over blind
• 8:28
  automation. Let them debate the architecture,
• 8:31
  but you must pull the final trigger. Let's take
• 8:33
  a moment to summarize this paradigm shift. We
• 8:35
  are moving from a world where they assign tasks
• 8:38
  to isolated tools. Now we are assigning roles
• 8:42
  to an integrated AI team. Jack Dorsey Buzz provides
• 8:46
  the open source room for that team to talk. They
• 8:49
  debate and refine work before it ever reaches
• 8:51
  human eyes, while letting us keep ownership of
• 8:53
  the resulting data. It changes how we scale our
• 8:56
  work completely. It does. And I want to leave
• 8:58
  you with something to chew on. Beat. If our AI
• 9:01
  agents are going to spend all day peer reviewing
• 9:03
  each other, catching mistakes and debating solutions
• 9:05
  in a shared room, at what point does the shared
• 9:08
  context of the AI team become more valuable than
• 9:11
  the institutional knowledge of the human company
• 9:13
  itself? Two -Sec Silence. Thank you for taking
• 9:16
  this deep dive with us. Outtiro Music.