AI Agent Architecture

Executive Summary

This section describes the Multi-Agent AI architecture for automating customization and deployment of ERP systems. The process is Spec-Driven: AI agents first generate strict technical specifications by comparing client requirements with the existing system analysis.

Section Contents

Chapter	Description
Knowledge Core	Knowledge base and inputs
Agent Roles	Roles of individual agents
Communication	Communication between agents
EspoCRM API	API endpoints for the Builder agent

---

Architecture Overview

sequenceDiagram
    participant User as Client
    participant KB as Knowledge Base
    participant A1 as Analyst Agent
    participant A2 as Architect Agent
    participant A3 as Builder Agent
    participant ERP as EspoCRM
    participant A4 as QA Agent

    User->>A1: Upload requirements
    A1->>KB: Query "Do we have module X?"
    KB-->>A1: "Yes, Inventory module; missing Feature Z"
    A1->>A2: Send Gap Analysis (Feature Z needed)

    Note over A2: SPEC-DRIVEN DEVELOPMENT
    A2->>A2: Write JSON Schema for Feature Z
    A2->>A2: Write UI Layout Spec
    A2->>A2: Write Gherkin User Stories
    A2->>A3: Hand off specifications

    A3->>ERP: Install base modules (Inventory)
    A3->>ERP: API create field "x_custom_field"
    A3->>ERP: API inject workflow script
    ERP-->>A3: Success (200 OK)

    A3->>A4: Start testing
    A4->>A4: Generate Playwright script from Gherkin
    A4->>ERP: Login & execute via browser
    ERP-->>A4: Rendered page
    A4->>A4: Verify field value
    A4->>User: Delivery (report + credentials)

---

Tooling Summary

Agent	Responsibility	Recommended tools
Librarian	Parse documentation, embeddings	LangChain, Pinecone, AST parsers
Analyst	Semantic search, gap detection	GPT-4o / Claude 3.5 Sonnet
Architect	Specifications, JSON/YAML	JSON Schema validation, Gherkin
Builder	Implementation via API/code	Python, EspoCRM API, CLI
QA Bot	Browser automation, testing	Playwright, Puppeteer

---

➡️ Continue to Knowledge Core.