Agentic Delivery Office
yaakovslonimczyk-sudo/agentic-delivery-office
Metodologia de entrega ágil multi-agent com staged gates e orquestração
Ver no GitHubKit de metodologia multi-agent para desenvolvimento de software com portões de qualidade e orquestração de sub-agentes especializados. Sistema portável que se instala uma vez no home directory e aplica-se a todos os projetos, com controle de gates humanos ordenados por custo de artefato e loops de review-and-fix determinísticos.
Instalação
npx convez-ai yaakovslonimczyk-sudo/agentic-delivery-office1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Orquestração de múltiplos agentes IA em pipelines de desenvolvimento
- Implementação de quality gates automáticos com revisão humana
- Padronização de processos de entrega em múltiplos projetos
Veredito

Sistema robusto e bem documentado para times que buscam automatizar entrega de software com múltiplos agentes mantendo controle humano em gates críticos. Exige experiência em CLI e padrões de desenvolvimento para implementação efetiva.
README
A tradução para o português deste README ainda está na fila.
The Agentic Delivery Office — Agent Methodology Kit
A portable, project-agnostic setup for multi-agent software development with minimal human intervention: specialist subagents with isolated contexts, staged human gates ordered by artifact cost, a bounded internal review-and-fix loop, and deterministic quality gates.
The methodology is explained in the article "The Agentic Delivery Office" (see article/MEDIUM-ARTICLE.md, with figures).
This is a system-level setup, not a per-project copy. You install it once into your home directory; the agents, protocol, Live Board, hooks, and slash commands then apply to every project. Per-project business context is added on demand with a slash command.
git clone https://github.com/yaakovslonimczyk-sudo/agentic-delivery-office.git
cd agentic-delivery-office
./install.sh # Claude Code, user scope (recommended)
Then, inside any project:
/office-init # adds this project's roster + business context
/office-board # opens the Live Board (blueprint org-chart / ⚽ match view)
Portable across providers (one methodology, three runtimes)
| Runtime | Command | Global drop targets |
|---|---|---|
| Claude Code | ./install.sh | ~/.claude/agents/*.md, ~/.claude/commands/, ~/.claude/settings.json, ~/.claude/office-board/ |
| OpenAI Codex | ./install.sh --target codex | ~/.codex/agents/*.toml, protocol appended to ~/.codex/AGENTS.md |
| Grok Build | ./install.sh --target grok | global skill ~/.grok/skills/agentic-office/ (Grok also auto-reads .claude/ + AGENTS.md) |
| legacy per-project | ./install.sh --scope project [PATH] | copies into PATH/.claude/ |
Add --dry-run to preview without changing anything. The installer is
idempotent and never clobbers — it merges hooks and appends marked blocks.
What's inside
agent-methodology-kit/
├── README.md ← you are here
├── install.sh ← copies the kit into any project
├── ORCHESTRATION.md ← the delivery-loop protocol (the methodology itself)
├── CLAUDE.md.template ← project-context template (fill in per project)
├── .claude/
│ ├── settings.json ← permissions allowlist + commit quality gate
│ ├── hooks/
│ │ └── gate-commit.sh ← blocks git commit unless lint + tests pass
│ └── agents/
│ ├── explorer.md ← read-only codebase mapper
│ ├── mockup-builder.md ← clickable HTML mockup for Gate 1
│ ├── planner.md ← Mode 1: HL component diagram · Mode 2: split build plan
│ ├── code-reviewer.md ← cold diff review, read-only
│ ├── test-writer.md ← tests from the SPEC, not the implementation
│ ├── security-reviewer.md
│ ├── db-reviewer.md ← schema/migration/EXPLAIN review
│ ├── design-reviewer.md ← UI critique with eyes (preview/browser tools)
│ └── business-reviewer.md ← one definition, invoked per lens (CFO/BizDev/BI/CEO)
└── ports/
├── PORTING.md ← how the concepts map across providers
├── codex/AGENTS.md ← drop at repo root for OpenAI Codex
└── grok/GROK-PROMPTS.md ← copy-paste prompt pack for Grok or any other LLM
Per-project setup (after the one-time system install)
The system install puts the agents and protocol in your home. The only thing
that is genuinely per-project is the business context (which grounds the
CFO/BizDev/BI/CEO review lenses) and the roster (the JUNIOR-tier line). Add
them by running /office-init inside a project — it interviews you for the
minimum context and writes the project's CLAUDE.md.
Scope — which projects the office runs in
A system-wide install does not impose the methodology on every repo. The
office is scoped via ~/.claude/office/scope.json, with a default policy:
opt-in(default, recommended for multi-project users): a project stays dormant — no gates, no commit gate, not on the board — until you run/office-initthere. New projects require explicit acceptance.opt-out: every project is active unless you exclude it.
Manage it with /office-scope (list, enroll, exclude, switch policy) or
/office-init scan to bulk-enroll repos under ~/dev. A --scope project
install auto-enrolls that project.
Prefer the legacy self-contained copy instead? ./install.sh --scope project
drops everything into the project's .claude/. In that mode, edit CLAUDE.md
(fill the placeholders — especially the
business context, which is what grounds the business-reviewer lenses), and run
a first feature through the loop described in ORCHESTRATION.md.
Design principles (why the kit looks like this)
- Agents are defined by tools + verifiable deliverable, not job titles.
Every agent either has restricted tools (reviewers can't edit) or a
deliverable that can be checked against ground truth (tests, screenshots,
EXPLAINoutput, fetched pricing data). - Fresh context is a feature. Reviewers see only the diff and the spec — never the implementer's reasoning — so they catch what the author can't.
- Silence is a valid review outcome. Every reviewer must return
APPROVEDwhen nothing rises to blocking. This is what keeps an autonomous fix loop from oscillating forever. - Blocking findings get adversarially verified before triggering rework. A refuter agent tries to kill each finding; only survivors get fixed.
- The loop is bounded. Max 2 fix-and-re-review rounds, then deliver to the human with residuals listed.
- Hooks enforce what prompts only suggest. The commit gate runs lint and tests mechanically; no agent can skip it.
- Staged gates, artifacts ordered by cost. Gate 1: clickable mockup + HL component diagram (cheapest direction test). Gate 2: full HLD + executive deck — produced ONLY after Gate 1 approval, since they elaborate the approved shape. Build & the review swarm run only after Gate 2. Deliberate shortcuts during the MVP slice are tracked in a list that the hardening phase must resolve — cuts never silently ship.
Office Board — live visibility
A zero-dependency local dashboard that shows the org chart in real time: which agent holds the ball right now (animated pass from the orchestrator), each agent's tier/model, cumulative tokens, run count, and an engagement log.
node .claude/board/server.js # from the project root
# open http://localhost:5599
How it works: the office-board.sh hook fires on every Agent-tool spawn
(PreToolUse) and return (PostToolUse — the return carries the subagent's
token usage), appending events to .claude/board/events.jsonl; the server
streams them to the browser over SSE. Hooks only see start/end of each
engagement, so token counters settle when an agent reports back — for
mid-run token streaming you'd graduate to Claude Code's OpenTelemetry
metrics. Agents outside the roster (e.g. ad-hoc general-purpose spawns)
appear automatically as amber "contractor" cards.
External agent dependency (Phase B)
The Design phase uses two user-level agents expected at ~/.claude/agents/:
hld-architect (full HLD with PlantUML diagrams) and deck-architect
(executive .pptx derived from the HLD). User-scoped agents are available in
every project on the machine, so the kit references them rather than bundling
them. On machines without them, use the generic HLD Writer / Deck Builder
role prompts in ports/ — the protocol is identical.
Token cost expectations
A full cycle (explore → plan → implement → parallel reviews → verify → fix → re-review) runs ~3–5× the tokens of single-agent work. The orchestrator's own context stays lean (it only ever sees summaries), which is what preserves quality over long autonomous sessions.
Model routing is policy, not improvisation — see "Model routing" in
ORCHESTRATION.md. The rule: tier ∝ how silent the agent's failure is. Tiers
are RELATIVE to the frontier, never hardcoded model names: SENIOR = the
session model (model: inherit — upgrades free when the frontier moves);
JUNIOR = one tier below, defined in exactly one line of the project's
CLAUDE.md roster and passed as a spawn-time override. Judging agents (planner,
code/security reviewers, refuters) are SENIOR and never downgraded; grounded
agents start JUNIOR and get promoted/demoted per project via evidence-driven
performance reviews logged in the roster. Junior-tier routing cuts ~30–40% of
spend at near-zero quality risk.