⚠️ DEPRECATED — this repo is archived

Consolidated into a single agents monorepo:

→ phant0um/claude-agent-dev — updated version lives in fullstack-agent-system/. /plugin installable, MIT.

Skills moved to phant0um/claude-skills.

---

---

tags: [fullstack-agent, sistema-multi-agente, claude, time-dev, engenharia] versao: "2.1.0" criado: 2026-05-14 atualizado: 2026-05-29

Fullstack Agent System

A senior engineering team composed of 6 specialized agents, orchestrated by a central Orchestrator. Each agent is a senior IT professional — not a tutor. Agents think, write compilable code, build integrations, and run validation tests.

Coordination via File-as-Bus: state lives in files, not in conversation history.
Mandatory Evidence: no deliverable without tests or execution logs.
Security with technical veto: can block any deploy.

---

Structure

fullstack-agent-system/
├── Orchestrator.md     ← Maestro: central planner (opus-4-8)
├── Backend-Dev.md      ← Stratum: APIs, DB, microservices (sonnet-4-6)
├── Frontend-Dev.md     ← Facet: UI/UX, React/Vue, a11y (sonnet-4-6)
├── Infra-Cloud.md      ← Bastion: AWS, Terraform, CI/CD (sonnet-4-6)
├── Data-AI.md          ← Neuron: ML, ETL, LLMs, RAG (opus-4-8)
├── Security.md         ← Sentinel: AppSec, OWASP, compliance (opus-4-8)
├── Forge.md            ← Forge: 5E code quality, scoring, refactoring (sonnet-4-6)
├── docs/
│   ├── Constitution.md     ← 6 principles that govern all agents
│   ├── agent-model-map.md  ← Routing: activity → model → agent
│   ├── standards-anti-patterns.md
│   └── progress.md         ← System state (File-as-Bus)
└── adr/
    └── 0000-template.md    ← Template for Architecture Decision Records

---

Model Map

Agent	opus-4-8	sonnet-4-6	haiku-4-5
Maestro	Complex planning	—	—
Stratum	—	APIs, DB, auth, refactoring	Tests, docs, seeds
Facet	—	Components, performance, a11y	CSS, E2E, Storybook
Bastion	—	IaC, CI/CD, observability	Dockerfiles, YAML, runbooks
Neuron	RAG architecture, ML design	ETL, analysis, features	Reports, docstrings
Sentinel	Threat modeling, compliance	Code review, IAM	Headers, configs, checklists
Forge	—	5E quality review, refactoring	Score reports, findings

Estimated savings: ~60–75% vs. using Opus for everything.

---

Usage in Claude Code

Specific project

# Activate Maestro as project CLAUDE.md
cp Orchestrator.md ./CLAUDE.md

# Specialist agent for a specific repository
cp Backend-Dev.md ./CLAUDE.md     # backend repo (Stratum)
cp Frontend-Dev.md ./CLAUDE.md    # frontend repo (Facet)
cp Infra-Cloud.md ./CLAUDE.md     # infra repo (Bastion)

Conditional rules by path

# Backend
.claude/rules/services.md      # paths: ["src/**/*.service.ts"]
.claude/rules/tests.md         # paths: ["**/*.test.ts", "**/*.spec.ts"]

# Frontend
.claude/rules/components.md    # paths: ["src/components/**/*.tsx"]
.claude/rules/styles.md        # paths: ["**/*.css", "**/*.scss"]

# Infra
.claude/rules/terraform.md     # paths: ["**/*.tf", "**/*.tfvars"]
.claude/rules/kubernetes.md    # paths: ["k8s/**/*.yaml"]

# Data
.claude/rules/dbt.md           # paths: ["dbt/**/*.sql", "dbt/**/*.yml"]
.claude/rules/ml.md            # paths: ["ml/**/*.py"]

# Security
.claude/rules/auth.md          # paths: ["src/auth/**", "**/middleware/**"]

Loading hierarchy in Claude Code

~/.claude/CLAUDE.md           ← Maestro (global, optional)
./CLAUDE.md                   ← Specialist agent (project)
.claude/rules/*.md            ← Conditional rules by path

---

Workflow

Maestro reads progress.md
  └─→ Breaks task into sub-tasks with done criterion
      ├─→ Stratum (APIs, DB, microservices)
      ├─→ Facet (UI, components, a11y)
      ├─→ Bastion (IaC, CI/CD, deploy)
      ├─→ Neuron (pipelines, ML, RAG)
      ├─→ Forge (5E quality review before merge)
      └─→ Sentinel (mandatory review on auth/data/infra)
          ↓
      Each specialist delivers:
        Deliverable + Evidence + State Update
          ↓
      Maestro validates Evidence, updates progress.md

---

Constitution (summary)

1. Evidence before code — spec exists before implementing; Evidence is mandatory
2. Security by default — Security has veto on any deploy
3. Tests are not optional — deliverable without tests = incomplete delivery
4. Closed scope by default — each agent does exactly what was delegated
5. Fail early, fail visibly — errors go to logs, never silent
6. System improves every cycle — hill-climbing via Vault SO (hill, review, guard)

---

System evolution

Based on the Nexus Agent System and the following sources:

• AiScientist (thin control + thick state, File-as-Bus)
• Agentic Harness Engineering (memory > prompts)
• Multi-Agent Orchestration patterns (Prep Line, Gordon Ramsay)
• RL Conductor (model routing by task type)

Previous version: Professor Fullstack (v1.0) — 4.5 models, no mandatory evidence, educational framing.