Jonathon's AI Wiki

Buzz (Jack Dorsey) — Agentic Slack on Swappable Harnesses

7 min read

Source: raw/Jack_Dorsey_s_Buzz_-_The_New_Hermes_Agent.md

Creator: Greg Isenberg (host) with guest Vinny (Wasp full-stack framework team) | Platform: YouTube | Feed: Greg Isenberg | URL: https://www.youtube.com/watch?v=_jGSgzBkzrY

Buzz is Jack Dorsey’s (Block) open-source “agentic Slack” — a team chat app where agents are first-class teammates rather than bolt-on integrations, sharing the full team context. Despite the video’s “New Hermes Agent” title, Buzz is not a Hermes product and not tied to any one agent runtime: the guest runs Claude Code as the harness under the hood, and Buzz ships adapters so any agent can swap in OpenClaw, OpenCode, or Goose — Hermes gets exactly one mention, as another harness you could adapt. The interesting design decision is that the harness and model under each agent are swappable without losing chat context.

Key Takeaways

  • Agents as teammates, not add-ons. Vinny: in Slack, agents arrive via integrations; in Buzz “they’re first-class citizens… members of your team.” Default agents ship out of the box; you can edit and add agents, and agent instructions are “basically like a system prompt.”
  • The harness is swappable under any agent. Vinny uses Claude Code exclusively as the harness, but Codex, OpenCode, and Goose are options, with adapters for others (OpenClaw, Hermes named). Switching harness or model carries all chat context to the new one — his answer to “model fatigue” from the breakneck release pace.
  • Pin agents to models by cost. His first setup move: Fizz pinned to a Fable model, Honey pinned to Sonnet, “because there are certain tasks that I don’t need to use the power of Fable and burn through tokens so fast.” This was his main setup advice.
  • A “Chief Agent Officer” delegation agent. As you accumulate specialized agents (copywriter, brainstormer, code reviewer), he created an agent whose job is to route: “I have this task. Who’s the best for it?”
  • Globally installed skills work. Skills and skill files in the global directory (not project-specific) are accessible to Buzz agents — “all the stuff under the surface that you’re used to using, if you use Claude Code… you don’t lose that stuff.”
  • Beta caveats are real. Workflows and recurring tasks “weren’t really landing great,” and Buzz feels slower than working directly in Claude Code or Codex because messages round-trip through your relay server. Vinny calls it beta, “might even be alpha.”
  • Sweet spot: solopreneurs and small teams (Greg frames it as 1–10 people, $0–10M revenue) iterating on small software and ideas — “if you’re really doing complex software engineering, this might not be the tool.”

What Buzz Does

  • Slack-shaped surface, agent-native mechanics: chat channels (private and public), audio huddles that agents can join with full chat context (Vinny hadn’t tested huddles himself), and agents that do end-to-end work from the conversation.
  • Git is tightly integrated. Agents create projects, cut feature branches, and work in parallel worktrees (copies — they don’t touch your local files). An experimental “projects” view (Settings → experiments: workflows, projects) shows the apps you’re building with agents.
  • Buzz hosts its own Git remotes on your relay, alongside optional GitHub integration. Relays are servers you host or Block hosts for you; all chats, data, and pushed code live there. Vinny reads this as Block “trying to take on more than just Slack… they’re trying to take on GitHub as well.”
  • End-to-end demo: he told an agent “spin up a simple CRM app using the Wasp full-stack framework and deploy it to Railway” — it built the app, pushed to remote repos, deployed live, and sent screenshots before he’d even checked what it was doing.
  • Closed-loop context workflows: he built a tweet-leaderboard full-stack app, exposed a public API from it, then set a Buzz workflow to check the API daily and post the stats into a channel — where he can reply and ask agent Fizz “what’s the common thread between my top tweets in the last 7 days?” (Sample output: his top two tweets cleared a million impressions combined, a weak theme collected ~88,000 — “novelty gets you reach, friction gets you engagement.”)
  • Prompting the agents: no special technique — “just have a conversation with them like a normal person”; they pull intent well because they hold all the channel context.

The Openness Thesis

  • Buzz is an open-source app built on Nostr, an open protocol. Vinny’s core selling point is openness, not the Slack-killer framing.
  • Data ownership: Slack controls your data and moving off is hard-to-impossible; same lock-in logic applies to models (“today it costs $200 a month for an unlimited plan, tomorrow it might be 2,000”). His argument: “you’re building your businesses on top of these tools, you should have flexibility and freedom and control.”
  • Shared compute: a setting suggests local models to download and share across the team — e.g., students pooling one beefy Mac Studio/Mac mini running a local model that everyone chats with from their own laptops.
  • Bitcoin Lightning is a plausible future: Nostr couples tightly with Lightning; Vinny’s guess (not shipped yet) is agents paying for compute/work and people tipping people for tasks inside Buzz.
  • Context as the foundation: Greg’s synthesis — Slack products have quietly become our context hubs, and models are only as good as the context you give them; Buzz treats the shared team context as the core and lets you pivot in any direction from it. Vinny ties this to why OpenClaw’s moment faded: “it was missing… this shared context, because we’re working in teams” — where a team can also just mean your other agents.
  • Community-in-the-loop: public channels can host your user community next to private team channels with shared context — a bug report in chat becomes “tell Fizz, let’s tackle this bug” or a prototype on the spot.

Try It

  • Start with the default agents; pin each agent to a model by task weight (frontier model for heavy agents, a cheaper model like Sonnet for routine ones) before anything else.
  • Keep your skills globally installed so Buzz agents inherit them across harness switches.
  • Treat workflows/recurring tasks as experimental for now; expect slower turnaround than direct Claude Code because of the relay round-trip.
  • Greg’s meta-advice (no affiliation with Buzz/Block/Dorsey, he notes): try tools like this to learn the shape of agent-native work — worst case you reconfigure Slack better; best case you’ve found the future of your team’s workflow.

Open Questions

  • Pricing and hosting terms for Block-hosted relays vs self-hosted ones — not covered in the episode.
  • How well audio huddles actually work — the guest had only seen third-party YouTube demos.
  • Whether Lightning payments (agent-pays-for-compute, tipping) ship, and in what form — currently the guest’s inference from the Nostr foundation, not an announced feature.
  • How Buzz’s per-agent harness adapters handle harness-specific features (hooks, MCP config) — not discussed.

Graph View

Backlinks