ECP AI Multi-Agent Squad
An AI agent squad for complete product development — from strategic context to production deploy — operated with Human in the Loop (HITL).
AI Multi-Agent Squad — The Concept
The ECP AI Multi-Agent Squad is a complete product squad operated by AI agents, where each traditional product team role (PM, Designer, PO, Architect, Devs, QA, Ops) is performed by a specialized agent.
The human acts as strategic decision-maker at 12 checkpoints (HITL gates) throughout the cycle, approving, rejecting or redirecting agent work. No agent advances without human approval.
The squad receives 3 markdown files + 1 target repository and generates all product artifacts: OKRs, backlog, prototypes, code, tests, CI/CD, dashboard and documentation.
Principles
The orchestrator coordinates and delegates. Each agent executes its specialty. The human decides.
Every output is recorded in JSON. Every human decision is documented. Every feedback is carried forward.
Each skill executes a single task and can be re-executed without side effects.
How It Works
briefing + tech + design
validate-inputs.mjs
coordinates phases
Strategy
Discovery
Delivery
Operations
Orchestrator Flow
validate-inputs.mjs) + behavioral validation of the 3 specs. If it fails, stops and reports.{REPO_DESTINO}/{phase}/.Required Inputs
The squad does not start without the 4 inputs below validated. The 3 markdown files must exist at the root of the target repository.
WHAT to build — functional scope, business rules, target audience and constraints.
Required sections:
- Product
- Problem
- Target audience
- Business context
- Core features
- Business rules
- Constraints
HOW to build — tech stack, architecture, patterns and inviolable rules.
Required sections:
- Tech stack
- Architecture
- Inviolable code rules
- Design patterns
- Deploy and infrastructure
- Tests
WITH WHAT LOOK — palette, typography, tokens, components and design guidelines.
Required sections:
- Color palette
- Typography
- Spacing and grid
- Components
- Theme / Mode
- Golden rule
WHERE to write — absolute or relative path of the folder where all squad outputs will be written.
{REPO_DESTINO}/
├── product_briefing_spec.md
├── tech_spec.md
└── design_spec.md
Two-Layer Validation
Cross-platform Node.js script that checks file existence and presence of required sections.
node validate-inputs.mjs {REPO_DESTINO}
The orchestrator reads the 3 files and verifies they have the required sections defined in shared/schemas/input-contracts.md.
Generates a JSON validation report.
Phase 01 — Strategic Context
What this phase produces
- Clear and inspiring product Objective
- Measurable and achievable Key Results tied to the objective
- Macro Opportunities per KR (Customer Pains, Needs and Desires)
- Leading indicators for each KR
- Product principles
- North Star Metric
What must NOT appear
- Features or functionalities
- Technical solutions
- Wireframes or visual references
- User stories
HITL #1 — Questions for the Human
- Is the Objective inspiring and does it guide product decisions?
- Are the KRs measurable and achievable within the defined timeframe?
- Are the Macro Opportunities linked to the right KRs?
- Are they real opportunities (pains/needs/desires) or disguised solutions?
- Is the scope appropriate for a discovery cycle?
Input: product_briefing_spec.md • Output: {REPO}/01-strategic-context/
Phase 02 — Product Discovery
6-Step Sequence
| Step | Agent | Action | HITL |
|---|---|---|---|
| 1. Prioritization | Product Owner | Receives the OST from Phase 01 and prioritizes opportunities by KR impact | #2 |
| 2. Ideation | Product Designer | Generates solution hypotheses for prioritized opportunities | #3 |
| 3. Low-Fi Prototype | Product Designer | Single HTML wireframe — mitigates Customer Value Risk | #4 |
| 4. High-Fi Prototype | Product Designer | Complete prototype with visual identity — mitigates Usability + Value Risk | #5 |
| 5. Structuring | Product Owner | Transforms approved hypotheses into Epics, Features and Stories (SAFe) | #6 |
| 6. 4 Risks Assessment | PM + Designer + PO | Collective assessment: Value (customer), Value (business), Usability, Feasibility |
Cagan’s 4 Risks
Do users want this?
Does it move the KRs?
Can users actually use it?
Can we build it?
Inputs: product_briefing_spec.md + design_spec.md • Output: {REPO}/02-product-discovery/
Phase 03 — Product Delivery
5-Step Sequence
| Step | Agent | Action | HITL |
|---|---|---|---|
| 1. Architecture | Software Architect | Bounded contexts, API contracts, DB schema, ADRs, repo scaffolding | #7 |
| 2. Back End | Back End Developer | Endpoints, persistence, services, authentication, tests | #8 |
| 3. AI Engineer | AI Engineer ∥ Front End | Chatbots, autonomous agents, RAG pipelines, AI guardrails (Claude API) | #8.5 |
| 3. Front End | Front End Developer ∥ AI Engineer | Components, pages, API and AI integration, fidelity to design_spec | #9 |
| 4. QA | QA | Shift-left, exploratory, automation, AI adversarial scenarios, quality gates | #10 |
Inputs: tech_spec.md + design_spec.md + approved backlog • Output: {REPO}/03-product-delivery/
Phase 04 — Product Operations
Sequence
| Step | Agent | Action | HITL |
|---|---|---|---|
| 1. Infrastructure | Operations | CI/CD pipeline, GMUD, installation guide, SLOs, DORA targets | #11 |
| 2. Dashboard + Docs | Operations | Operational dashboard (4 sections) + documentation site (all phases) | #12 |
Feedback Loop — HITL #12 Decisions
| Result | Action |
|---|---|
| Variation wins and moves KRs | Full rollout. PM updates baseline. Next cycle. |
| No variation moves KRs | Return to Phase 02 — revisit opportunities and hypotheses. |
| KRs worsen | Return to Phase 01 — revise OKRs and OST. |
| Technical bug | Return to Phase 03 — fix and new A/B round. |
Input: tech_spec.md • Output: {REPO}/04-product-operation/ + {REPO}/docs/
Orchestrator
The orchestrator never executes tasks directly. Its role is to coordinate, delegate, manage HITLs and ensure the flow across all 4 phases.
Responsibilities
- Receive the request and identify which phase the work is in
- Trigger the correct agent with full context
- Present outputs to the human at HITL checkpoints
- Carry human feedback to the correct agent
- Manage feedback loop when A/B indicates need
- Keep
/shared/memory/context.jsonupdated
Orchestrator Rules
- Execute a skill directly
- Advance without HITL approval
- Assume — always carry full context
- Start without the 4 validated inputs
- Include previously approved outputs in the next agent’s context
- Document human feedback before returning to agent
- On error, try 1x with additional context before escalating
- Inform the agent of the
{REPO_DESTINO}path
Squad Agents
10 specialized agents (1 orchestrator + 9 operational), each with subagents and atomic skills. Total: 38 subagent groups and 138 skills.
Subagents & Skills
Each agent has specialized subagents that execute atomic skills. Full catalog below.
Product Manager — Subagents & Skills
Product Designer — Subagents & Skills
Product Owner — Subagents & Skills
Software Architect — Subagents & Skills
Back End Developer — Subagents & Skills
AI Engineer — Subagents & Skills
Front End Developer — Subagents & Skills
QA — Subagents & Skills
Operations — Subagents & Skills
HITL Gates — Human in the Loop
12 checkpoints where the human evaluates, approves, rejects or redirects agent work. No agent advances without explicit approval.
| HITL | Phase | Post-Agent | Core Question |
|---|---|---|---|
| #1 | 01 | Product Manager | OKRs correct? OST without solutions? |
| #2 | 02 | PO Prioritization | Right opportunities prioritized? |
| #3 | 02 | Designer Ideation | Hypotheses approved for prototyping? |
| #4 | 02 | Designer Low-Fi | Concept validated? Value risk mitigated? |
| #5 | 02 | Designer High-Fi | Usability validated? Visual identity correct? |
| #6 | 02 | PO + PM + Designer | 4 risks approved? Ready for Delivery? |
| #7 | 03 | Software Architect | Solid architecture? Contracts ready? |
| #8 | 03 | Back End Developer | APIs correct? Persistence validated? |
| #8.5 | 03 | AI Engineer | AI solutions integrated? Guardrails active against prompt injection and leakage? |
| #9 | 03 | Front End Developer | Front integrated with backend and AI? Design faithful to design_spec? |
| #10 | 03 | QA | Quality approved? Ready to operate? |
| #11 | 04 | Operations | Environment ready? Deploy authorized? |
| #12 | 04 | A/B Testing | Winning variation? What did we learn? |
Possible decisions at each HITL
Advances to the next agent or phase.
Advances with documented restrictions.
Returns to agent with detailed feedback.
Communication Contracts
All agents communicate via standardized JSON. The orchestrator carries context between agents.
{
"phase": "01-strategic-context",
"agent": "product-manager",
"mission": "mission description",
"context": {},
"inputs_from_previous": {},
"hitl_feedback": "feedback (if return)"
}
{
"status": "success | error | partial",
"phase": "current phase",
"agent": "agent-name",
"subagents_used": [],
"skills_used": [],
"deliverables": {},
"open_questions": [],
"next_hitl": 1
}
{
"hitl": 1,
"decision": "approved | rejected",
"feedback": "human comments",
"next_agent": "next agent"
}
{
"status": "error",
"agent": "agent-name",
"error_code": "INVALID_INPUT",
"message": "description",
"retry_suggested": true
}
How to Run
Prerequisites
- Claude Code installed and configured
- Node.js 20+ (for the validation script)
- The 3 markdown files created in the target repository
Step by Step
product_briefing_spec.md, tech_spec.md and design_spec.md at the root of the target repository following the template in shared/schemas/input-contracts.md.node validate-inputs.mjs /path/to/target-repo
If it returns exit code 1, fix the errors before proceeding.
cd ecp-ai-multi-agent-squad
claude
Start Phase 01 — Strategic Context.
Target repository: /path/to/target-repo
The orchestrator will run the pre-flight check, present the validation report and, if approved, trigger the Product Manager Agent.
approved— advance to the next agentapproved with caveats— advance with restrictionsrejected+ feedback — return to agent with instructions
docs/ and the operational dashboard.Squad Folder Structure
ecp-ai-multi-agent-squad/
├── CLAUDE.md # Entry contract + global rules
├── validate-inputs.mjs # Pre-flight check cross-platform (Node.js)
├── README.md
├── agents/
│ ├── orchestrator/ # Coordination and HITL
│ │ └── CLAUDE.md
│ ├── product-manager/ # Strategy and metrics
│ │ ├── CLAUDE.md
│ │ └── subagents/
│ │ ├── market-research/ (4 skills) competitor-analysis, customer-segmentation, market-sizing, trend-scanning
│ │ ├── product-vision/ (4 skills) vision-writer, north-star-definer, product-principles, opportunity-brief-writer
│ │ ├── product-strategy/ (4 skills) okr-facilitator, four-risks-assessor, assumption-mapper, outcome-roadmap
│ │ └── metrics/ (4 skills) experiment-designer, leading-lagging-indicators, ab-test-analyzer, report-generator
│ ├── product-designer/ # Research, ideation and prototypes
│ │ ├── CLAUDE.md
│ │ └── subagents/
│ │ ├── customer-research/ (4 skills) continuous-interview-partner, ethnographic-researcher, assumption-tester, insight-synthesizer
│ │ ├── ux/ (8 skills) low-fi-prototyper, high-fi-prototyper, ideation-facilitator, solution-sketcher, information-architecture, interaction-design, accessibility-checker, usability-evaluator
│ │ └── ui/ (4 skills) component-spec, design-tokens, responsive-guidelines, visual-consistency-checker
│ ├── product-owner/ # Prioritization and structuring
│ │ ├── CLAUDE.md
│ │ └── subagents/
│ │ ├── epic/ (4 skills) epic-writer, enabler-epic-writer, opportunity-brief-writer, opportunity-prioritizer
│ │ ├── feature/ (2 skills) feature-writer, enabler-feature-writer
│ │ └── story/ (4 skills) user-story-writer, acceptance-criteria-writer, story-splitter, functional-documentation-writer
│ ├── software-architect/ # Domain and architecture
│ │ ├── CLAUDE.md
│ │ └── subagents/
│ │ ├── domain-design/ (6 skills) bounded-context-mapper, aggregate-designer, class-model-designer, ubiquitous-language-builder, context-map-writer, class-diagram-generator
│ │ ├── system-design/ (5 skills) adr-writer, api-contract-designer, data-modeling, evolutionary-architecture, repo-scaffolder
│ │ ├── evaluation/ (4 skills) build-vs-buy, risk-assessment, technical-debt-tracker, tech-radar
│ │ └── enablement/ (4 skills) architecture-pairing, dependency-audit, pattern-advisor, scalability-analysis
│ ├── backend-developer/ # API and persistence
│ │ ├── CLAUDE.md
│ │ └── subagents/
│ │ ├── api/ (4 skills) api-builder, auth-implementer, contract-enforcer, error-handler
│ │ ├── persistence/ (4 skills) schema-builder, repository-implementer, migration-manager, query-optimizer
│ │ ├── quality/ (4 skills) unit-test-writer, api-test-writer, integration-test-writer, mock-builder
│ │ └── living-docs/ (4 skills) code-commenter, api-documenter, schema-documenter, adr-logger
│ ├── ai-engineer/ # Chatbots, AI agents, RAG and guardrails
│ │ ├── CLAUDE.md
│ │ └── subagents/
│ │ ├── chatbot-builder/ (4 skills) conversation-designer, claude-api-integrator, context-manager, prompt-engineer
│ │ ├── agent-builder/ (4 skills) mcp-server-builder, tool-use-designer, agent-orchestrator, autonomous-workflow
│ │ ├── rag-pipeline/ (4 skills) document-processor, embedding-pipeline, retrieval-optimizer, grounded-response
│ │ └── ai-guardrails/ (4 skills) prompt-injection-shield, data-leakage-preventer, scope-enforcer, safety-tester
│ ├── frontend-developer/ # Interface and integration
│ │ ├── CLAUDE.md
│ │ └── subagents/
│ │ ├── implementation/ (5 skills) component-builder, design-to-code, responsive-implementer, state-management, mobile-identity
│ │ ├── quality/ (4 skills) unit-test-writer, integration-test-writer, e2e-script-writer, accessibility-implementer
│ │ ├── performance/ (4 skills) core-web-vitals-monitor, rendering-strategist, bundle-optimizer, caching-strategist
│ │ └── living-docs/ (4 skills) code-commenter, component-documenter, adr-logger, readme-writer
│ ├── qa/ # Quality and tests
│ │ ├── CLAUDE.md
│ │ └── subagents/
│ │ ├── shift-left/ (4 skills) story-quality-reviewer, testability-advisor, risk-identifier, three-amigos-facilitator
│ │ ├── exploratory/ (4 skills) charter-designer, session-based-tester, heuristic-tester, edge-case-finder
│ │ ├── automation/ (4 skills) api-test-writer, e2e-script-writer, test-data-builder, performance-tester
│ │ └── quality-intelligence/ (4 skills) bug-reporter, postmortem-writer, blameless-retrospective, quality-metrics
│ └── operations/ # CI/CD, infra and A/B
│ ├── CLAUDE.md
│ └── subagents/
│ ├── flow/ (7 skills) pipeline-builder, deployment-strategist, feature-flags-manager, dora-metrics-tracker, toil-reducer, docs-generator, installation-guide-writer
│ ├── infrastructure/ (4 skills) iac-writer, capacity-planner, environment-manager, finops-optimizer
│ ├── reliability/ (4 skills) monitoring-setup, slo-sla-designer, incident-commander, blameless-postmortem-writer
│ ├── devsecops/ (4 skills) vulnerability-scanner, compliance-as-code, secrets-manager, security-pipeline-integrator
│ ├── dashboard/ (3 skills) sre-metrics-designer, product-metrics-designer, dashboard-builder
│ └── change-manager/ (3 skills) gmud-writer, risk-assessor, change-approver
├── phases/
│ ├── 01-strategic-context/ PHASE.md + README.md
│ ├── 02-product-discovery/ PHASE.md + README.md
│ ├── 03-product-delivery/ PHASE.md + README.md
│ └── 04-product-operation/ PHASE.md + README.md
└── shared/
├── identity/ Template for design_spec.md
├── memory/ context.json (session state)
└── schemas/ Communication and input contracts
Generated Outputs in Target Repository
{REPO_DESTINO}/
├── product_briefing_spec.md [INPUT]
├── tech_spec.md [INPUT]
├── design_spec.md [INPUT]
│
├── 01-strategic-context/ OKRs, OST, vision, principles
│ ├── phase-01-output.json
│ └── hitl-1.json
│
├── 02-product-discovery/ Opportunities, backlog, prototypes
│ ├── prototype/
│ │ ├── low-fi.html Navigable wireframe
│ │ └── high-fi.html Complete prototype with visual identity
│ ├── epics/
│ ├── features/
│ ├── stories/
│ └── hitl-2..6.json
│
├── 03-product-delivery/ Architecture, ADRs, code, tests
│ └── hitl-7..10.json
│
├── 04-product-operation/ SLOs, A/B, DORA, CI/CD, GMUD
│ ├── ops-output.json
│ ├── dashboard-output.json
│ └── hitl-11..12.json
│
└── docs/ Navigable documentation site
├── index.html
├── fase-01..04.html
├── assets/
├── gmud/
├── dashboard/
└── INSTALACAO.md
Agent Bibliographic References
- Teresa Torres — Opportunity Solution Tree, Continuous Discovery
- Marty Cagan — 4 Product Risks, Empowered Teams
- Bernard Marr — OKRs and KPIs
- SAFe 6.0 — Epic > Feature > Story hierarchy
- IDEO — Design Thinking
- Don Norman — Usability
- Eric Evans — Domain-Driven Design
- Martin Fowler — Evolutionary Architecture
- Sam Newman — Building APIs
- Kent Beck — TDD / BDD
- Gene Kim — DevOps Handbook
- Google — SRE Book
- Nicole Forsgren — Accelerate / DORA
- Anthropic — Claude API, Messages API, Tool Use, MCP
- OWASP — LLM Top 10 (prompt injection, data leakage)
- Lewis et al. — Retrieval-Augmented Generation (RAG)
- Yao et al. — ReAct (Reasoning + Acting agents)
- OpenAI / Anthropic — Best Practices for System Prompts
Changelog v3.3
Update from March 25, 2026 — new AI Engineer agent with 4 subagents and 16 skills for AI solutions integrated into the product.
| Item | v3.2 | v3.3 | What changed |
|---|---|---|---|
| Agents | 9 | 10 | +AI Engineer (Phase 03, parallel with Front End) |
| Subagent groups | 34 | 38 | +chatbot-builder, +agent-builder, +rag-pipeline, +ai-guardrails |
| Skills | 122 | 138 | +16 AI skills (4 per subagent) |
| HITL Gates | #1–#12 | #1–#12 + #8.5 | +HITL #8.5 post-AI Engineer (guardrails and AI integration) |
| Phase 03 | 4 sequential steps | 5 steps (AI Eng ∥ Front) | AI Engineer and Front End run in parallel after Back End |
| QA | Functional tests | +AI adversarial scenarios | QA now includes prompt injection, leakage and scope tests |
AI Engineer — Details
Designs, implements and protects AI solutions integrated into the product using the Anthropic Claude API as the primary backend.
Required Security Principles:
- Input Sanitization — sanitize all input before sending to the LLM
- System Prompt Protection — system prompts never exposed to the end user
- Output Validation — validate LLM responses before returning
- Scope Enforcement — reject questions outside the domain
- PII Detection — block personal data leakage
- Rate Limiting — limit requests per user/IP
- Logging & Audit — log interactions for auditing (no PII)
- Fallback Graceful — safe responses when the LLM fails
Integration with other agents:
- Software Architect → API contracts for AI endpoints and ADRs
- Backend Developer → HTTP routes that AI Engineer consumes/exposes
- Frontend Developer → chat/conversation components with the AI API
- QA → AI adversarial scenarios beyond functional tests
- Operations → API costs, LLM latency and AI-specific SLOs
Changelog v3.2 (Mar 17, 2026)
| Item | v1.1 | v3.2 | What changed |
|---|---|---|---|
| Agents | 8 | 9 | Orchestrator included as official agent |
| Subagent groups | 30+ | 34 | Exact count of all groups |
| Skills | 70+ | 122 | Complete inventory of all skills |
| PO / Story | 3 skills | 4 skills | +functional-documentation-writer |
| Architect / Domain Design | 4 skills | 6 skills | +class-model-designer, +class-diagram-generator |
| Frontend / Implementation | 4 skills | 5 skills | +mobile-identity |
| Folder tree | Skill counts | Counts + names | Each group now lists all skill names |