A founder-level teardown of Block's new agent-chat app — the auto-approval default, the identity-key risk, and why the real question isn't which AI agent wins.
Buzz's real news isn't whether it beats Hermes Agent or OpenClaw — Block ships auto-approved commands, a single impersonation-capable identity key, and no read-only agent role by default, so the founder-level question is whether an agent's work this month survives you closing the laptop and stays auditable by your team.
Buzz is Block's new open-source chat app where AI agents sit in channels as named members instead of being a bolted-on chatbot — and buried in its own source code is a list of 12 agent programs, including rivals Hermes Agent and OpenClaw, that plug in underneath it via the open Agent Client Protocol. The real story isn't which agent wins: Buzz auto-approves agent commands by default, hands out a single identity key that can impersonate the whole account, has no read-only agent role, and grades its own approval gates as 'glue still drying.' It's free to run, but agent usage still bills to your own API key, and Block just cut concurrent agents from 24 to 10. Treat it as a room your existing agent works in, lock channel membership down to yourself, and keep a named human approving anything before it reaches a client.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →
Cold open: a list of agent programs buried in Block's own code, including Hermes Agent and OpenClaw, that almost no coverage of Buzz has actually opened.
Reframes 'does Buzz beat the agent I already pay for' as the wrong question — the real one is whether this month's agent work survives closing the laptop and stays visible to teammates.
Discloses no payment from Block, no contact with anyone there, and a commercial interest in Hermes Agent, a Buzz-adjacent competitor.
Quick channel intro: builds AI-powered businesses and automations, teaches creators and founders to use AI tools.
Opening Buzz doesn't drop you into a chat — it drops you onto a key screen first, which matters for the identity-key discussion later.
Buzz is a chat app where the agent is a member of the channel, not a bolted-on feature — it has its own name in the member list and stays in the room whether you're typing or not. Announced July 21, free and open source, self-hostable.
Pulls the actual harness-description file from Block's public repo: 12 named agent programs, 4 built directly into Buzz and 8 shipped as presets, including Hermes Agent and OpenClaw.
The open standard Buzz and its plugged-in agents use to talk to each other, started at Zed in 2025 and since spun into its own project — nothing locks a user's agent to Block.
Block ships its own built-in agent inside Buzz, competing at the same level as any plugged-in agent — Buzz gives your existing agent somewhere to work rather than replacing it.
Demo: switching the default harness from Claude Code to Goose took two clicks with no export step and no migration; Buzz auto-installed the Claude Code adapter itself in about 20 seconds.
The two names the biggest videos put in their titles as rivals — Hermes and OpenClaw — are, to Block, just entries in a dropdown; Buzz is a fraction of their size and about a week old with a releases page moving daily.
Testing whether a conversation survives an agent swap needs a live channel with history, which lives on a hosted relay via BuilderLab — the first thing in Buzz that isn't free the way open source trains you to expect.
Sets up the reveal: one line in Block's published code decides whether a human ever sees a request before it runs — and it's why the creator changed a setting before letting Buzz do anything.
Block's own code comment says permission requests are auto-approved with an 'allow once' option — the default, not something a user left switched on. Also flags the five-question go-or-no-go checklist from the creator's paid community.
Anyone who can post in a channel can make its agent run commands on your computer. Buzz hands new users an identity key with a warning: never share it, anyone holding it can impersonate you and reach your whole account.
Block filed an open issue against itself: a hidden instruction in a message could talk the agent into handing over the identity key. The issue was still open and a proposed fix was closed without merging.
Proposes a minimal, auditable workflow: one channel with one job — an enquiry in, the agent drafts against pinned documents, a named human checks it, one output goes out.
Nobody would let a machine near a client quote unencumbered — which is exactly why a named human belongs in the room to check the draft before anything leaves.
Projects and Workflows both ship switched off by default, hidden behind an experiments panel — easy to spend an evening hunting for a menu that was never live.
Block's own README grades every feature by how finished it is, and workflow approval gates land in the half-finished row — their words are 'glue still drying.' The thumbs-up gets recorded but doesn't stop the agent.
The app is free but the models aren't — the Codex path wants an OpenAI API key rather than a subscription, so agent usage bills to the user's own key, not Block's.
Block cut concurrent agents from 24 to 10, and a busy channel spends money with nobody typing since agents replay every unanswered mention on startup. There's no read-only agent role, and the activity log is tamper-evident but not tamper-resistant.
Reality-check list: nothing pages you when a run stops, Buzz needs an always-on machine (agents stop the moment it sleeps), and phone-only users can't even get started since the phone just holds the key.
People already paying for Hermes Agent or OpenClaw are probably the best fit right now — their harness stays where it is, and Buzz just adds a room around it.
Channel membership is the whole permission model, so client data needs real consideration before pointing Buzz at client work — none of this is a prescription, check the current release yourself.
Closing challenge: take the last quote or invoice an agent handled and ask whether a teammate could see what got decided and who approved it — that's the only reason this video exists.
Before you point any agent-chat app at real work, check whether it auto-approves commands, whether one credential can impersonate your whole account, and whether a human actually has to click before anything ships.
“There's a list inside Jack Dorsey's new app that, in my opinion, almost nobody covering it has opened.”
“Nothing stops to ask you before the command runs.”
“A room full of robots talking to each other isn't a business.”
“Their words for it are glue still drying.”
“Tamper-evident, but not tamper-resistant.”
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
Everyone covering Buzz is asking whether it beats the AI agent you already pay for. Johnny Nel opens Block's own source code instead and argues that's the wrong question — the one that matters is whether the work your agent did this month survives you closing the laptop, and whether anyone else on your team could see who approved it.
An open standard for how a chat surface like Buzz talks to whatever agent program is plugged in underneath it — started at Zed in 2025, now an independent project. Buzz speaks the protocol instead of making private deals with each agent vendor.
A minimal, auditable agent workflow: one channel, one job, with a named human approver gating anything before it reaches a client.
A pre-install checklist the creator runs before letting any agent tool near a machine that matters, meant to save a wasted weekend on a tool that was never going to fit. Full list is gated behind his paid community and not disclosed in the video.
“It's waiting for you inside the AI founder's vault along with the rest of the guys.”
Soft, repeated pitch for his paid Skool community (AI Founders Vault) woven through the video as on-screen cards rather than one hard sell at the end — first at the 'five-question go-or-no-go' beat (~6:01) and again near the close (~10:59–11:34) alongside a subscribe ask.
00:01
00:0800:14
00:20
00:2900:35
00:40
00:44
00:4700:50
01:03
01:1101:2101:29
01:35
01:41
01:48
01:5402:05
02:20
02:34
02:4803:02
03:16
03:2603:34
03:41
03:51
04:0304:15
04:2804:4204:56
05:09
05:2105:33
05:42
05:50
05:57
06:0706:21
06:34
06:4506:54
07:02
07:1007:1707:23
07:31
07:4107:50
07:5808:05
08:11
08:1708:23
08:28
08:3408:41
08:47
08:55
09:0509:14
09:22
09:29
09:3509:43
09:53
10:0310:13
10:2110:28
10:3510:40
10:51
10:57
11:04
11:14
11:20
11:29Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.
14:16An agency owner walks through the workflows where an AI agent now drafts YouTube packaging, brand vision docs, recruiting outreach and social carousels, and where a human still has to step in.
September 16th
24:49AI Edge's host installs seven free, high-star GitHub repos live on screen, turning stock Claude Code into a social-research engine, a browser operator, an ads department, and a 279-agent company.
September 16th
27:30Greg Isenberg and Remy put the invite-only Instinct agent through real errands, a haircut, a restaurant table, and a Bali visa, and find out it keeps a copy of your email even after you disconnect it.
September 15th
19:28A software architect sketches phantom-fs, the abstraction layer that lets dozens of Claude Code agents share one synced, conflict-free codebase from anywhere.
September 16th
15:58A 16-minute walkthrough of replacing a $99+/month ManyChat plan with a free, self-hosted Instagram DM automation, installed end to end by a Claude Code agent skill.
September 11th
08:22A working editor sets up DaVinci Resolve's new Claude MCP integration, gives it one real editing task, and races it on the clock.
September 10th