GenAI Architecture

Appearance

Master Plan ​

Canonical Source of Trust: Google Doc Tab Master
Training Program: [Huy Chau] Generative AI Training Plan

www.agilityio.com

Generative AI Training Plan ​

Sep** 16, 2026**

OVERVIEW ​

This training plan details the development of AI agents and multi-agent applications. It covers core agent concepts, practical implementations, and real-world applications, progressing from foundational understanding to sophisticated agentic systems. The curriculum emphasizes hands-on learning through projects, focusing on practical software engineering approaches to building LLM-powered applications.

PREREQUISITES ​

  • Python/TypeScript programming skills and experience
  • Basic understanding of web APIs and data structures
  • Jupyter notebook environment setup

TIMELINE ​

  • FUNDAMENTAL PLAN: 8 Weeks
  • **BEYOND PLAN: **2 Weeks

KEY OBJECTIVES ​

  • Master Context Engineering: Moving beyond "writing prompts" to systematically managing the entire information ecosystem (knowledge, memory, and environment).
  • Implement Harness Engineering: Designing "Build-Verify-Fix" loops and scaffolding to mold "inherently spiky" model intelligence for production reliability.
  • Develop Stateful Orchestration: Using LangGraph and its Functional API to build durable, multi-agent systems with persistence and human-in-the-loop capabilities.
  • Production-Grade Security & Evaluation: Applying the OWASP Top 10 for Agentic Applications (2026) and using LLM-as-a-judge for trajectory-based evaluation.

EXTRA NOTES ​

  • No forces on the programming language, but primarily choosing between TypeScript and Python, depending on your base.
  • You are expected to be proactive and productive in leveraging appropriate AI tools for learning and practice to ensure the timeline remains on track.

FUNDAMENTAL PLAN ​

PHASE 1: THE ENGINE & PROGRAMMATIC PROMPTING (1 Week) ​

  • Introduction to GenAI & "Spiky Intelligence"
  • Engine Parameters
  • System Prompt Design
  • Instrumentation & Tracing Hierarchy
  • Compulsory Security
  • NIST AML Taxonomy

PHASE 2: THE ENGINEERING OF REASONING (1 Week) ​

  • Logical Decomposition
  • Strategic Exploration
  • Abstraction Techniques
  • Interactive Reasoning
  • Trajectory Analysis
  • Compulsory Security

PHASE 3: MODULAR PIPELINES & STRUCTURE DATA (1 Week) ​

  • The Runnable protocol & LCEL (Pipes & Blocks)
  • Standardized Content Blocks
  • Structured Outputs via Pydantic
  • Pipeline Instrumentation
  • Tracing Hierarchy
  • Compulsory Security: Content Moderation & Secure Schema Enforcement

PHASE 4: CONTEXT ENGINEERING & AGENTIC RAG (1.5 Weeks) ​

  • The 3-Layer Memory (Working, Episodic, Semantic)
  • Context Window Management
  • RAG Evolution (Agentic, GraphRAG)
  • Question Transformation (HyDE, RRF, Multi-Query)
  • The RAG Triad & Agent Reliability Metrics
  • Evaluation Dataset Design
  • Offline Evaluation
  • Compulsory Security

PHASE 5: HARNESS ENGINEERING & MCP CONSUMPTION (1.5 Weeks) ​

  • The AI Harness & Scaffolding
  • Model Context Protocol (MCP) tool integration
  • Build-Verify-Fix (BVF) Loops
  • Deterministic Middleware (Loop Detection & Pre-Completion Checklists)
  • Trace-Driven Failure Analysis
  • The Data Flywheel (Manual Prototype)
  • AIRateLimiter & Operational Resource Controls

PHASE 6: STATEFUL GRAPHS & HUMAN-IN-THE-LOOP (2 Weeks) ​

  • Stateful Orchestration (LangGraph 1.0 & DCGs)
  • Strongly Typed Shared State (TypedDict)
  • Durable Persistence & Checkpointing
  • Human-in-the-Loop (HITL) & Native Interrupts
  • The Functional API (@entrypoint, @task)
  • Tracing Hierarchy
  • Online Evaluation
  • Evaluation Lifecycle
  • Security: Privileged Action Gates & Audit Trails

PRACTICE: NEXUS ​

Follow this specification to practice

BEYOND PLAN ​

Reference

  • Automated Governance & Secure Infrastructure
  • The Internet of Agents (IoA) — A2A & Layer 9
  • Autonomous Coordination & Shared Memory
  • Industrial Evaluation, Meta-Optimization & Token Economics
  • Real-Time Data Streaming & Headless Orchestration
  • Google ADK (Optional) Doing practice for those topics by following this specification

PRODUCTION EVALUATION ENGINEERING ​

Reference

  • LLM-as-a-Jugde at Scale
  • Annotation Queues for Expert Calibration
  • The Data Flywheel (Automated)
  • Scalable Supervision

ENTERPRISE-GRADE GENAI SECURITY PIPELINE ​

Reference

  • 15-layer security pipeline design
  • GenAI defenses:
    • Zero-Trust Dual-Service model
    • Quad-Layer Guardrails
    • Deterministic Tool Security
    • Context Sanitization

EXPANSION ​

LangChain ecosystem frameworks by use-cases:

Following this plan for learning Deep Agents rapidly in 1 week.

ENGINES & TOOLS ​

  • Orchestration: LangGraph v1 (Stateful Graph & Functional API)
  • Evaluation: Promptfoo, RAGAS, DeepEval
  • Observability: LangFuse (Tracing, Trajectory Evals, Polly AI Assistant), LangSmith
  • Security: LangChain Guardrails, AWS Bedrocks Guardrails, LLM Guard, LlamaFirewall
  • Optimization: Redis (Semantic Caching), vLLM (Inference)
  • Production-Grade Code Structure: Template

REFERENCES ​

  • Book:** **AI Agents and Applications
  • Context Engineering Guide: RAG, Memory Systems & Dynamic Context (2026)
  • Harness Engineering for Agentic Coding Systems (LangChain Case Study)
  • OWASP Top 10 for LLM (2025) and Agentic Applications (2026)
  • LangGraph v1 Documentation & Functional API Guide
  • Security Best Practices

Master AI Architecture Training Program

Huy Chau © 2026