⬡ ECP
Experiment 804 — AI Multi-Agent Squad

The Experiment

1 Context

In 2025, many solutions emerged related to agentic AI.

  • Claude Code
  • Devin
  • Codex
  • OpenClaw
  • LangGraph

2 Problem

Evaluate the Impact that this could have on the entire product development cycle (Strategic Context, Product Discovery, Product Delivery, Product Operations) — not just on the software development stage — and consequently on our competitiveness.

  • Speed
  • Productivity
  • Efficiency
  • Quality
  • Security

3 Solution Hypothesis

Build a squad of connected autonomous AI agents to develop and evolve digital products.

Constraint: fast process with minimal investment of hours.

Solution: AI Multi-Agent Squad

Specialized Agents

🧑‍💼 Product Manager
🎨 Product Designer
📋 Product Owner
🏛️ Software Architect
🔧 Back End Dev
🤖 AI Engineer
💻 Front End Dev
🧪 QA Engineer
⚙️ Operations
🎯 Orchestrator

Squad Structure

10
Agents
36
Sub-agents
141
Skills
12
HITLs / product

4 Product Management Aspects

01

Strategic Context

OKRs, North Star, product vision, market

02

Product Discovery

Research, ideation, prototypes, backlog

03

Product Delivery

Architecture, backend, frontend, QA

04

Operations

CI/CD, GMUD, deploy, SLOs, A/B

Full Squad documentation →

What was built: Digital product ecosystem

🏦

ECP Bank (Personal)

  • Send and receive Pix with daily and overnight limits
  • Virtual card with statement and purchases
  • Smart statement with categorization
  • Multi-agent AI chatbot (Claude)
  • Dashboard with cash flow and notifications
Access → Documentation →
🏢

ECP Emps (Business)

  • Business account with corporate Pix and CNPJ keys
  • Bill issuance with barcode and QR Code
  • Corporate cards with per-holder limit
  • Team management with RBAC (admin/finance/viewer)
  • Personal↔Business transfers within same ownership
Access → Documentation →
💳

ECP Pay

  • Pix, credit card and registered bill
  • Provider Pattern: internal mode (simulated) and Asaas (real)
  • Automatic payment split between parties
  • Token vault (card vault) cross-app
  • Operational admin panel with dashboard and audit log
Access → Documentation →
🍕

ECP Food

  • Restaurant catalog with search and filters
  • Multi-restaurant cart with free shipping ≥ R$ 120
  • 3 payment methods (ECP Card, Credit, Pix QR)
  • Pix payment with real-time SSE and webhooks
  • Restaurant panel + platform admin
Access → Documentation →

How the products connect

🏦
ECP Bank Personal
bank.ecportilho.com
🏢
ECP Bank Business
emps.ecportilho.com
↕
↕
🍕
ECP Food
food.ecportilho.com
↔
💳
ECP Pay
Central payment hub
↕
↕
🔄
Internal Gateway
Local simulation
🏧
Asaas External Gateway
Real production

Integrated flow: order in FoodFlow

  • User selects food and places an order in ECP Food
  • Chooses a payment method in ECP Food
  • User confirms the order
  • ECP Food calls the ECP Pay payment gateway
  • ECP Pay interacts with the personal bank to add an item to the credit card statement
  • ECP Pay interacts with the business bank to credit the purchase amount to the restaurant

Results

Complexity

1.002
Business Complexity Points
57
Features
33.082
Lines of Code
4
Products
17.6
Avg BCP / feature

Productivity

~635x
Acceleration Factor
~34s / BCP
Time / BCP (AI)
~9h 28min
AI Squad Effort
6h / BCP
Reference / BCP (human)
6.012h
Estimated Human Squad Effort

Speed (Lead Time)

~28h 16min
Total Lead Time with HITL
~17h 16min
Lead Time without HITL
~11h
HITL Pause (review)
~47
Total HITLs
~3.56M
Tokens consumed
~US$ 139
Total cost (Claude Opus)
~R$ 805
Total cost in BRL
R$ 0.80
Cost per BCP
View details by product →

v2 count (code-only): BCP calculated using source code only as evidence. Reference: 6 hours per BCP — CI&T v1.5 benchmark for a traditional human team. Pricing: Claude Opus 4 (input $15/1M + output $75/1M). Exchange rate: R$ 5.80.