# buzz-block-panorama-deep-research-2026-08-12

## Veille

Internal research report dated **August 12, 2026** consolidating, for presentation purposes, everything publicly documented about **Buzz** — **Block**'s humans + agents workspace, launched on **July 21, 2026** under the **Apache 2.0** license. It aggregates the two engineering posts already filed alongside the corporate announcement, the GitHub repository, press coverage, X, and **three independent hands-on accounts** that constitute the dossier's only non-self-reported data. **(A) A vocabulary gap documented by quotation**: **Jack Dorsey**'s launch tweet announces *"model-agnostic, decentralized, self-sovereign, and open source"*; Block's `ARCHITECTURE.md` states *"The relay is the single source of truth. All reads and writes flow through it. There is no peer-to-peer event exchange, no gossip, no replication."* The relay is therefore single and authoritative per community: Buzz's "decentralization" is an **organizational sovereignty** — self-hosting and portable identity — not network redundancy. **TFTC**'s formulation: *"Two of those three hold cleanly. The third needs a qualifier."* **(B) An asymmetry between demonstrated rigor and exploitation risk.** On one side, a rare degree of formalism for a v0.4.x/0.5.x: multi-tenant isolation specification **mechanized in TLA+**, authorization properties verified in **Tamarin**, a model-checked Git storage protocol, a hash-chained append-only audit log, 127 *event kinds*, NIP-01/42/98/34. On the other, channel membership is the unit of permission — *"channel membership is not fine-grained tool authorization"* (João Queirós) —, agents run in `--dangerously-skip-permissions` outside any sandbox on a human's machine, and observability is lacking: *"Buzz tells me an agent got a message. It doesn't tell me what happens next"* (DevTools Daily, which reports silent OOM kills). Block acknowledges it: *"the agent can do anything, and security rests entirely on restricting who can tell it what to do"*. **(C) The technical stack**, absent from the filed posts: **Rust** relay (Axum WS + REST), **Postgres**, **Redis**, **S3/MinIO** via Blossom, **Tauri + React** desktop client. Agent integration goes through **`buzz-acp`**, an **ACP** harness that plugs in goose, Codex and Claude Code and translates **ACP ↔ MCP**, plus **`buzz-agent`**, an in-house agent. The report corrects itself on one point: the *"+33% more work"* in Block's TL;DR is the **ratio of completed tasks (20 versus 15 out of 44)**, not a score gain — the score itself rises from 59.1% to 71.5%, i.e. **+12.4 points**.

## Titre Article

Buzz (buzz.xyz) — Rapport de recherche pour présentation

## Date

2026-08-12

## URL

*Aucune URL publique — rapport de recherche interne non publié au moment de la mise en fiche. Source archivée dans `raw-data/buzz-block-panorama-deep-research-2026-08-12.md`.*

## Keywords

Buzz, buzz.xyz, Block, Jack Dorsey, agentic workspace, humans and agents, channel, Apache 2.0, organizational sovereignty, decentralization, single relay, single point of failure, no replication, ARCHITECTURE.md, Nostr, NIP-01, NIP-42, NIP-98, NIP-34, event kinds, keypair, cryptographic identity, portable identity, delegated authorization, audit log, append-only, hash chain, tamper-evidence, tamper-resistance, compromised relay operator, end-to-end encryption, Rust, Axum, Postgres, Redis, MinIO, Blossom, Tauri, React, Git on object storage, TLA+, Tamarin, formal verification, multi-tenant isolation, Agent Client Protocol, ACP, buzz-acp, buzz-agent, MCP, ACP MCP translation, composability via protocols, goose, Codex, Claude Code, OpenRouter, vLLM, llama.cpp, Ollama, Hive, Swarm, Terminal-Bench 2.1, Long-Horizon Terminal-Bench, dangerously-skip-permissions, channel-based authorization, agent observability, silent OOM, mandatory attribution, mobile app, QR pairing, EXIF, NIP-PL, BuilderBot, Agentic AI Foundation, Linux Foundation, AGENTS.md, DevTools Daily, TFTC, Sundar Pichai, Justin Waldron, multiplayer agent harness, v0.x maturity, caveats

## Authors

**Deep Research Veille Interne** — rapport non signé, produit le **12 août 2026** en préparation d'une présentation. Aucune URL publique ; source archivée dans `raw-data/`.

Rapport de veille et non source primaire : ce qu'il apporte est la couche que Block ne publie pas — l'`ARCHITECTURE.md` du dépôt, les métriques GitHub, la presse, et **trois évaluations indépendantes** (João Queirós, DevTools Daily, darrenjrobinson). Chaque affirmation porte sa source, les verbatims sont donnés en anglais avec traduction, et les URL X sont distinguées selon qu'elles ont été vérifiées, vérifiées par contenu seulement, ou non retrouvées. Une section **Caveats** finale liste six réserves, dont *« Les benchmarks proviennent de Block (auto-évaluation), sur son propre setup. »* Le document est explicitement outillé pour une présentation (section Recommandations, ordre des slides) : à lire comme un livrable de préparation.

## Ton

**Profile**: structured research report (TL;DR → Key Findings → details across eight sections → Recommendations → Synthesis → Caveats), analyst register, in French with English verbatims systematically translated. Audience: whoever must present Buzz to peers without getting caught out.

**Style**: the dominant device is **juxtaposition** rather than judgment — the report places the marketing formulation next to the technical documentation and lets the gap speak for itself, Dorsey's tweet against the `ARCHITECTURE.md`. Traceability serves as the argument: full URLs, dates, GitHub issue numbers (#2484, #2367), acquisition amounts (Slack → Salesforce **$27.7B** on 07/21/2021; GitHub → Microsoft **$7.5B** on 06/04/2018), diverging version numbers by source. The document is built to withstand a hostile question in a meeting, and it embraces its own rhetorical recommendation: *"An honest 'limitations' slide… strengthens the credibility of the presentation."*

**Marker phrases borrowed from its sources**:
- ***"The bottleneck moved from intelligence to coordination"*** (Longwell)
- ***"authorization does not erase authorship"*** (Block)
- ***"channel membership is not fine-grained tool authorization"*** (Queirós)
- ***"Buzz tells me an agent got a message. It doesn't tell me what happens next"*** (DevTools Daily)
- ***"Two of those three hold cleanly. The third needs a qualifier"*** (TFTC)
- ***"It's the first proper multiplayer agent harness"*** (Waldron)
- ***"rough edges and giant chasms"*** (Block)

**Epistemic stance**: bounded and audited — what is verified is stated as such, what is not is named as such. Acknowledged blind spot: no enterprise adoption data, no hosting price, no retention figures, since Block does not publish any. The report translates this into a re-evaluation criterion rather than speculation.

## Pense-betes

- **Date / source**: **August 12, 2026**, unpublished internal watch report. Aggregates Block posts, `ARCHITECTURE.md`, press, X and three independent accounts.
- **Key framing**: launch vocabulary is not architecture documentation. Transposable method: for any platform that calls itself decentralized, read `ARCHITECTURE.md` **before** the announcement post. ### The decentralization / architecture gap | Source | Statement | |---|---| | **Jack Dorsey** tweet | *"model-agnostic, decentralized, self-sovereign, and open source"* | | Block's `ARCHITECTURE.md` | *"The relay is the single source of truth. All reads and writes flow through it. There is no peer-to-peer event exchange, no gossip, no replication."* | | **TFTC** | *"Two of those three hold cleanly. The third needs a qualifier."* | A single relay per community constitutes a single point of failure. The sovereignty on offer is organizational — self-host, carry your identity with you — not network redundancy. ### Demonstrated rigor vs exploitation risk | What is formally verified | What is not | |---|---| | multi-tenant isolation **mechanized in TLA+** | the unit of permission is **channel membership** | | authorization properties verified in **Tamarin** | agents run in `--dangerously-skip-permissions`, outside any sandbox | | model-checked Git storage protocol | agent observability is lacking — silent OOM kills, abandoned tasks appearing active | | hash-chained append-only audit log | — | The formal proof covers identity and storage; the exploitation risk sits in tool authorization and visibility. Don't let the former reassure on the latter. ### A security distinction worth keeping ***Tamper-evidence ≠ tamper-resistance.*** Signed events prove who did what; they do not prevent a compromised relay operator from **deleting** events. On the relay hosted by Block, messages are not end-to-end encrypted, and model providers may receive prompts and channel content. Question to ask any vendor selling a "tamper-proof" audit log: tamper-proof against modification, or against deletion? ### Figure correction Block's TL;DR announces *"33% more work"*: this is the ratio of **completed tasks — 20 versus 15 out of 44** —, not a score gain. The score rises from **59.1% to 71.5%**, i.e. **+12.4 points**, of which 11.4 come from the additional completions. The report itself juxtaposes the two in its synthesis, an ambiguous phrasing not to be copied verbatim into a slide. Full detail, including the negative result on short tasks, in [[patel-block-buzz-teams-tokens-benchmarks-2026-08-06]]. ### Technical stack and deployment **Rust** relay (Axum WS + REST), **Postgres** (events + full-text search), **Redis** (pub/sub, presence, typing), **S3/MinIO** (media via Blossom), **Tauri + React** desktop. **127 event kinds**; NIP-01 (feed), NIP-42 (login auth), NIP-98 (REST auth), NIP-34 (Git events). Deployment via a production Docker Compose (Postgres, Redis, MinIO, Caddy/TLS), one-click Railway, or Block's free hosted relay in beta. Rust 1.88+, Node 24+, pnpm. ### Interoperability **`buzz-acp`** is an **Agent Client Protocol** harness (born at Zed, JSON-RPC 2.0 over stdio, now under community governance) that plugs in **goose**, **Codex** and **Claude Code**, and translates **ACP ↔ MCP** — ACP on the client/agent side, MCP on the tools side. Added to this is **`buzz-agent`**, an in-house agent: up to **8 concurrent sessions**, each with its own MCP servers, history and context; Anthropic Messages API, OpenRouter, or any OpenAI-compatible endpoint (vLLM, llama.cpp, Databricks, Ollama). A formulation worth keeping: *"The agent does not know what MCP server it talks to. The MCP server does not know what agent is calling it. They compose through protocols, not imports."* For the acronym trap around ACP, see [[girard-acp-deux-protocoles-un-sigle-2026-08-02]]. ### Genealogy **goose** (Block, **January 28, 2025**, Apache 2.0, CTO Dhanji Prasanna) is the ancestor, contributed in **December 2025** to the Linux Foundation's **Agentic AI Foundation** alongside MCP and AGENTS.md — 27,000+ stars, 350+ contributors. **BuilderBot**, Block's first internal Slack agent, provides the original problem: shared credentials, model switching, identity management. Block has been building a position on the open agentic layer for eighteen months. ### Market stakes **Bradley Axen** (Head of AI Capabilities, Block): *"Every company is going to need a place where humans and agents work together. The question is whether that place is proprietary or open."* **Justin Waldron** (Zynga co-founder): *"Buzz is not a slack killer. It's much bigger. It's the first proper multiplayer agent harness."* Reception: ~25,900 GitHub stars (~7,600 in three days), a Dorsey tweet at ~2.3–2.7M views, TechCrunch / SiliconANGLE / Decrypt coverage, public endorsement from Sundar Pichai. ### Independent field reports | Source | Contribution | |---|---| | **DevTools Daily** (~07/26) | two-person workspace, self-hosted on a VPS behind Caddy, with Codex, Copilot and Claude agents — successful consolidation, real agent-to-agent handoffs within the same thread; observability is missing | | **João Queirós** (07/23) | first in-depth evaluation; origin of the channel-authorization formulation | | **darrenjrobinson** | three agents with distinct identities in a self-hosted Buzz, each with its own keypair and its own billing — *"qualitatively different from a bot integration"* | Various guides: agents only act on @mention; without a stop instruction, two agents can loop into infinite replies; the hosted onboarding has had bugs (issues #2484, #2367) making self-hosting more reliable at this stage. ### Mobile app (Tom Brow, 07/29) The app does not host agents: it **signs messages** and connects directly to relays; identity comes from the desktop via **QR pairing**, with no standalone identity path on mobile. Privacy: no analytics SDK, **EXIF metadata stripped** before upload, push following a **NIP-PL** draft where relays cannot see the device token and the gateway sees neither public keys, content, metadata, nor relay identity. A reference model for designing push without metadata leakage. ### What's missing for production use No published price for managed hosting, no enterprise adoption figures, no fine-grained per-tool authorization, no agent observability, unsigned Windows binaries, scattered docs, an early-stage Git forge UI. Block states it plainly: *"rough edges and giant chasms"*. The three proposed re-evaluation thresholds: evaluate self-hosted on a non-critical project if the team is already *agent-heavy*; wait for fine-grained per-tool authorization and observability before any production use; re-evaluate once a hosting price and a v1.0 are published. ### How to use this fiche It is the **context dossier**; primary sources lie elsewhere. For the thesis and identity model → [[longwell-block-buzz-workspace-agents-nostr-2026-07-21]]. The dossier ages fast: releases every few days, GitHub metrics in motion, statuses dated mid-August 2026. Worth noting: to help people leave GitHub, Buzz hosts its own code on GitHub.

## RésuméDe400mots

Internal research report dated **August 12, 2026** consolidating the public state of **Buzz**, **Block**'s humans+agents workspace launched on **July 21, 2026** under **Apache 2.0**, for presentation purposes. It aggregates Block's two engineering posts, the corporate announcement, the GitHub repository, press coverage, X and **three independent evaluations** — this last layer carrying most of the added value.

**The concept.** Buzz merges team chat, a Git forge and automated workflows into a single space where agents are **full members, not bots**. The thesis is Tyler Longwell's: *"The bottleneck moved from intelligence to coordination."* Bradley Axen (Head of AI Capabilities) frames the market stakes: *"Every company is going to need a place where humans and agents work together. The question is whether that place is proprietary or open."*

**The architecture.** A **Rust** relay on **Nostr** (NIP-01/42/98/34, 127 *event kinds*), **Postgres**, **Redis**, **S3/MinIO**, **Tauri+React** desktop. Each participant holds a keypair; every message, review, workflow step and Git event is **signed** into a hash-chained append-only audit log. A rare degree of formalism for a **v0.4.x/0.5.x**: multi-tenant isolation mechanized in **TLA+**, authorization properties verified in **Tamarin**. Agent integration goes through **`buzz-acp`**, an **ACP** harness that plugs in goose, Codex and Claude Code and **translates ACP ↔ MCP** — *"They compose through protocols, not imports."*

**The central gap.** Jack Dorsey announces *"decentralized, self-sovereign"*; Block's `ARCHITECTURE.md` states: *"The relay is the single source of truth… There is no peer-to-peer event exchange, no gossip, no replication."* A single relay per community, hence a **single point of failure**: decentralization is **organizational sovereignty**, not redundancy.

**The limitations, documented.** The unit of permission is **channel membership** — *"channel membership is not fine-grained tool authorization"*; agents run in **`--dangerously-skip-permissions`**, outside any sandbox; **observability is lacking** (*"It doesn't tell me what happens next"*, silent OOM kills). Signed events are *tamper-evident*, not *tamper-resistant*: a compromised relay operator can delete them. On the hosted relay, there is **no end-to-end encryption**.

**A figure correction.** The "+33% more work" is the **ratio of completed tasks (20 vs 15 out of 44)**, not a score gain — which rises from 59.1% to 71.5%, i.e. **+12.4 pts**.

**Reception**: ~25,900 GitHub stars, a Dorsey tweet at ~2.3-2.7M views, endorsement from Sundar Pichai, and Justin Waldron's formulation: *"the first proper multiplayer agent harness"*. Acknowledged caveats: benchmarks **self-evaluated by Block**, no published hosting price, no adoption figures.

## GrapheDeConnaissance

- Block —publie→ Buzz (TECHNOLOGIE, 0.98)
- Buzz —est_instance_de→ workspace collaboratif humains-agents auto-hébergeable (CONCEPT, 0.96)
- Jack Dorsey —affirme_que→ Buzz est model-agnostic, décentralisé, self-sovereign et open source, et vise à réduire la dépendance de Block à Slack et GitHub (CITATION, 0.95)
- Buzz —s_oppose_à→ la qualification de décentralisé au sens réseau : l'architecture officielle pose le relais comme source unique de vérité, sans échange pair-à-pair, sans gossip et sans réplication (AFFIRMATION, 0.96)
- Buzz —permet→ une souveraineté organisationnelle par auto-hébergement et identité portable, et non une redondance réseau (AFFIRMATION, 0.94)
- relais unique —s_oppose_à→ la disponibilité du workspace, dont il constitue le point de défaillance unique (AFFIRMATION, 0.93)
- Buzz —utilise→ Nostr (TECHNOLOGIE, 0.97)
- Buzz —utilise→ un relais écrit en Rust avec Postgres pour les événements et la recherche, Redis pour la présence et le pub/sub, et un stockage objet S3 ou MinIO pour les médias (AFFIRMATION, 0.94)
- Block —utilise→ TLA+ (TECHNOLOGIE, 0.94)
- Block —utilise→ Tamarin (TECHNOLOGIE, 0.92)
- Tamarin —permet→ de vérifier les propriétés d'autorisation de Buzz, ce qui est inhabituel pour un logiciel encore en version 0.5.x (AFFIRMATION, 0.9)
- Buzz —s_oppose_à→ l'autorisation fine par outil : l'appartenance au canal reste l'unité de permission, un agent dans un canal pouvant faire ce que font les membres (AFFIRMATION, 0.95)
- João Queirós —affirme_que→ l'appartenance à un canal n'est pas une autorisation fine par outil (CITATION, 0.94)
- Buzz —utilise→ un lancement en --dangerously-skip-permissions, hors bac à sable, sur la machine d'un humain (AFFIRMATION, 0.93)
- Block —affirme_que→ l'agent peut tout faire, et la sécurité repose entièrement sur la restriction de qui peut lui donner des instructions (CITATION, 0.94)
- Buzz —s_oppose_à→ l'observabilité de l'activité des agents : le workspace signale qu'un agent a reçu un message mais pas ce qu'il en fait, des agents ayant subi des kills OOM silencieux (AFFIRMATION, 0.92)
- journal d'audit signé —permet→ de prouver qui a fait quoi, sans empêcher un opérateur de relais compromis de supprimer des événements (AFFIRMATION, 0.93)
- relais hébergé par Block —s_oppose_à→ le chiffrement de bout en bout des messages, les fournisseurs de modèles pouvant recevoir prompts et contenus de canal (AFFIRMATION, 0.92)
- buzz-acp —utilise→ Agent Client Protocol (TECHNOLOGIE, 0.95)
- buzz-acp —permet→ de brancher goose, Codex et Claude Code dans Buzz en traduisant entre ACP côté agent et MCP côté outils (AFFIRMATION, 0.94)
- buzz-agent —utilise→ Model Context Protocol (TECHNOLOGIE, 0.93)
- buzz-agent —permet→ jusqu'à huit sessions concurrentes disposant chacune de ses serveurs MCP, de son historique et de son contexte, contre l'API Anthropic Messages, OpenRouter ou tout endpoint compatible OpenAI (AFFIRMATION, 0.92)
- Block —affirme_que→ l'agent ignore à quel serveur MCP il parle et le serveur MCP ignore quel agent l'appelle : ils se composent par des protocoles, pas par des imports (CITATION, 0.94)
- goose —est_basé_sur→ le programme open source de Block lancé le 28 janvier 2025, ancêtre conceptuel de Buzz (AFFIRMATION, 0.92)
- goose —fait_partie_de→ Agentic AI Foundation (ORGANISATION, 0.93)
- Bradley Axen —affirme_que→ toute entreprise aura besoin d'un lieu où humains et agents travaillent ensemble, et la question est de savoir si ce lieu est propriétaire ou ouvert (CITATION, 0.95)
- Bradley Axen —travaille_chez→ Block (ORGANISATION, 0.95)
- Buzz —concurrence→ Slack (TECHNOLOGIE, 0.9)
- Buzz —concurrence→ GitHub (TECHNOLOGIE, 0.88)
- Justin Waldron —affirme_que→ Buzz n'est pas un tueur de Slack mais le premier véritable harnais d'agents multijoueur (CITATION, 0.9)
- Deep Research Veille Interne —mesure→ le gain de 33 % annoncé par Block correspond au ratio de tâches terminées, vingt contre quinze sur quarante-quatre, et non au score qui progresse de 59,1 % à 71,5 % (MESURE, 0.93)
- Deep Research Veille Interne —affirme_que→ les benchmarks disponibles sur Buzz proviennent d'une auto-évaluation de Block sur son propre dispositif, sans mesure indépendante (AFFIRMATION, 0.94)
- Buzz —s_oppose_à→ un usage en production tant que l'autorisation fine par outil et l'observabilité des agents ne sont pas livrées (AFFIRMATION, 0.9)

---
Canonical: https://www.thekb.eu/en/fiches/buzz-block-panorama-deep-research-2026-08-12/