Skip to content

6SNT-RADIO/6SNT-Studio

Repository files navigation

6SNT-Studio — a gated, 10-agent software studio for Claude Code (agent-teams)

6SNT-Studio

A gated, 10-agent software studio for Claude Code (agent-teams).

English · Español

License: MIT Built with Claude Code Mode: agent-teams Agents: 10 Human gates: 6 Entry: greenfield + brownfield Privacy: local-first

Ten specialized AI agents build software in a disciplined pipeline. Each agent owns exactly one deliverable, and nothing advances without explicit human approval at its gate. The result is AI-built software with the predictability of a real engineering process — not a one-shot prompt.

At a glance: 10 agents · 6 human gates · 11 principles · 8 execution rules · 3-tier models (opus/sonnet/haiku) · 2 entry paths (greenfield + brownfield) · local-first.

The studio is not a project — it is the infrastructure that serves every project. Projects come and go; the studio stays.

— Built and operated by CA6SNT (Valdivia, Chile). Method-first, low-bureaucracy: the procedure is the product.


🚀 Install (Claude Code)

In a Claude Code session:

/plugin marketplace add 6SNT-RADIO/6SNT-Studio
/plugin install studio@6SNT-Studio

That installs the 10 agents, the skills, the hooks, the /startsnt and /adopt commands, and the project scaffolder. Private repo? run gh auth login once first.

Prerequisite — enable agent-teams. The studio runs on Claude Code's experimental agent-teams. A plugin can't enable this for you, so add the flag to your ~/.claude/settings.json and restart Claude Code:

{ "env": { "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1" } }

(Optional: OpenTelemetry env vars for local observability — see docs/PIPELINE.md.)

Then, in a session rooted in your project folder: new idea? run /startsnt (greenfield, guided intake). Existing project? run /adopt (brownfield on-ramp).


🧭 The idea in one picture

flowchart TD
  A01([01 · Strategist]):::opus --> G01{{G01 · BRIEF}}:::gate
  G01 --> A02([02 · Researcher]):::sonnet --> G02{{G02 · RESEARCH}}:::gate
  G02 --> A03([03 · UX/UI]):::sonnet --> G03{{G03 · IDENTITY}}:::gate
  G03 --> A04([04 · Architect]):::opus --> G04{{G04 · ARCHITECTURE}}:::gate
  G04 --> A05([05 · Data Modeler]):::opus
  A05 --> A06([06 · Frontend]):::sonnet
  A05 --> A07([07 · Backend]):::sonnet
  A06 --> A08([08 · QA]):::sonnet
  A07 --> A08
  A08 --> G08{{G08 · QA}}:::gate
  G08 --> A09([09 · Security]):::opus --> A10([10 · Tech Writer]):::haiku --> G10{{G10 · RELEASE}}:::gate

  classDef opus fill:#8957e5,stroke:#5a2ea6,color:#fff;
  classDef sonnet fill:#2ea44f,stroke:#1b6e34,color:#fff;
  classDef haiku fill:#d97757,stroke:#a64b32,color:#fff;
  classDef gate fill:#f0b429,stroke:#b35900,color:#1a1300;
Loading

Boxes are colored by model tier (opus / sonnet / haiku); the amber hexagons are the human gates (PO approval). 06 ∥ 07 run in parallel.

A human Product Owner (PO) approves every gate. A Lead orchestrates (creates tasks, assigns owners, closes gates) but never writes code or deliverables. The ten agents do the work — one artifact each.

Two ways in

Greenfield — an idea → /startsnt → the pipeline above. Brownfield — an existing codebase (built elsewhere/by hand) → /adopt: an on-ramp that grounds the real code, reconstructs its specs (marking inferences [TO VERIFY]), passes one human gate, then rejoins the pipeline at the project's real maturity. Not a second pipeline — an on-ramp.

flowchart LR
  P([existing project]) --> GR[ground + secret scan]:::g --> SM[smoke: does it run?]:::g --> RC[reconstruct specs·TO VERIFY]:::g --> GA{{G-ADOPT}}:::gate --> RJ[rejoin pipeline at real maturity]:::g
  classDef g fill:#2ea44f,stroke:#1b6e34,color:#fff;
  classDef gate fill:#f0b429,stroke:#b35900,color:#1a1300;
Loading

🧩 Three pieces

Piece Role
Lead / Orchestrator Coordinates: classifies work size, builds the task graph, routes escalations, closes gates. Produces no deliverables.
Agents 01–10 Each executes its stage when activated and delivers a single artifact (see AGENTS).
Product Owner (human) Approves deliverables at each gate and makes the calls agents can't. No agent replaces the PO.

✅ Why it works

  • One agent, one deliverable. A hard ownership map (see AGENTS) — an agent may read anything but only writes its own zone. Boundary clashes are escalated, never resolved between agents.
  • Proportional gating. Work is classified TRIVIAL / STANDARD / COMPLEX; only the gates that matter run. Adding gates to a gated pipeline increases failure — so the studio defaults to the smallest size that fits.
  • Definition-of-Done is mechanical. "Done" includes an artifact that starts, proven by an automated smoke test. Release is blocked until artifact + smoke-pass + assets exist.
  • Adversarial, cost-aware review. An external critic attacks each deliverable before its gate (external review beats self-review); subjective gates (brand) use a keyless multi-tier jury, not a single judge.
  • Enforcement where it's deterministic, norms where it can't be. Session-level guards are real locks; per-agent ownership is a norm audited post-hoc via OpenTelemetry traces. See PIPELINE.

📚 Documentation

Doc What's inside
docs/AGENTS.md The 10 agents: role, model tier, single deliverable, ownership, escalation topology.
docs/SKILLS.md The studio's skills & mechanisms: critic, smoke test, brand rubric + jury, intake gating, brownfield grounding, scaffolding, evals, observability.
docs/PIPELINE.md The two entry paths, the 6 human gates + G-ADOPT, the task graph, size classification, and what's mechanical vs. norm.
docs/PRINCIPLES.md Principles P-01…P-11 and execution rules RC-01…RC-08 (the studio's constitution).

The runnable pieces live at the repo root: agents/ (the 10 agents), skills/ (the studio skills, incl. critic and adopt-ground), hooks/, commands/ (/startsnt, /adopt), scripts/ (the scaffolder) and templates/. Spanish reference copies of the agents are under i18n/es/.


🎚️ Model tiers (cost-aware)

Tier Agents
opus 01 Strategist · 04 Architect · 05 Data Modeler · 09 Security
sonnet 02 Researcher · 03 UX/UI · 06 Frontend · 07 Backend · 08 QA
haiku 10 Technical Writer

The lead may bump a tier for a one-off hard task.


📦 Status

  • ✅ Method validated by building real desktop apps end-to-end through the full gated pipeline (QA caught real domain bugs; Security remediated findings; brand & architecture passed machine rubrics).
  • ✅ Runs on Claude Code agent-teams with native OpenTelemetry observability and promptfoo evals as a pre-gate.
  • 🌐 Active agents/skills/commands are English; Spanish reference copies live under i18n/es/.

🔧 Adapting it to your own work

The studio is project-agnostic by design (principle P-10): no agent hard-codes a specific project. To adopt it, you keep the constitution (PRINCIPLES) and the ownership map (AGENTS), then point the agents at your stack. The reference stack here is Electron + TypeScript desktop apps, but the method is stack-neutral.


⚖️ License

MIT © 2026 CA6SNT (Luis Soto). Use it, adapt it, build with it.

About

CA6SNT Agent Studio - a gated 10-agent pipeline for Claude Code (agent-teams). Two entry paths: greenfield (/startsnt) + brownfield (/adopt). Specs, principles & method.

Topics

Resources

License

Contributing

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors