Advanced Agentic AI Workshop: From AI User to AI Architect


Instructor	Nat Weerawan (ณัฐ วีระวรรณ์)
Version	2.0
Created	13 January 2026
Contact	[email protected]

Executive Summary

Workshop สอนการสร้าง AI System ที่ทำงานแทนคุณได้ — ไม่ใช่แค่ตอบคำถาม

🎯 Key Deliverables

Deliverable	Description
1. Personal AI Agent	Agent ที่ customize แล้ว รู้ style และ context ของคุณ
2. Personal Oracle	Second Brain ที่จำทุกอย่างได้ — ใช้งานได้จริง
3. Multi-Agent Skills	ทักษะการสั่ง 5-10 agents ทำงานพร้อมกัน
4. Production Patterns	วิธีการ deploy และ maintain ใน enterprise

⚠️ ข้อควรระวัง

Multi-agent ทรงพลัง แต่ต้องเข้าใจก่อน
ถ้าใช้ไม่เป็น อาจสร้างความสับสน
Workshop = sandbox ปลอดภัย → ฝึกก่อนใช้จริง

🎯 Who Is This For?

Audience	Fit	Why
Developers ที่ใช้ ChatGPT ทุกวัน	⭐⭐⭐⭐⭐	พร้อม level up เป็น agent architect
Tech leads building AI products	⭐⭐⭐⭐⭐	ต้องเข้าใจ patterns ก่อน scale
DevOps/Platform engineers	⭐⭐⭐⭐	Automation + AI = powerful combo
Product managers (technical)	⭐⭐⭐	เข้าใจ capabilities, ไม่ต้อง code เยอะ
Executives (awareness)	⭐⭐	Consider Executive Overview option

Part 1: Course Options

Option A: Executive Overview (Half Day)

Best for: Decision makers, CTO/CIO, non-technical stakeholders

Time	Module	Content	Type
09:00-09:45	The AI Agent Revolution	ChatGPT → Claude → Agents → Multi-Agent	Lecture
10:00-11:00	Live Demo	ดู multi-agent system ทำงานจริง	Demo
11:15-12:00	Business Cases	ROI, use cases, implementation roadmap	Discussion
12:00-12:30	Q&A + Next Steps	Planning your AI journey	Consultation

Price: ฿60,000 (up to 30 participants)

Deliverable: Understanding + implementation roadmap

💡 Insight: Half-day เหมาะสำหรับ executives ที่ต้องการ understand ก่อน invest — ไม่ต้อง hands-on แต่ได้เห็นของจริง และรู้ว่า team ต้อง learn อะไร

Option B: Fundamentals (1 Day)

Best for: Teams starting their AI agent journey

Time	Module	Content	Type
09:00-10:00	AI Evolution	AI → Agent → Multi-Agent concepts	Lecture
10:15-11:15	Claude Code Setup	Install, configure, first agent chat	Hands-on
11:30-12:30	CLAUDE.md Mastery	Create your agent's identity	Hands-on
13:30-14:30	Subagent Basics	Spawn agents, delegate tasks	Learn by Doing
14:45-15:45	Skills & Hooks	Extend your agent's capabilities	Hands-on
16:00-17:00	Live Build	Build something real together	Workshop

Price: ฿60,000 (up to 30 participants)

Deliverable: Personal AI agent with custom skills

💡 Insight: 1 วันพอสำหรับ "aha moment" — ผู้เข้าอบรมจะกลับบ้านพร้อม agent ที่ทำงานได้จริง และรู้ว่าจะ extend ยังไง แต่ยังไม่มี memory system

Option C: Professional (2 Days) ⭐ Recommended

Best for: Teams building production AI systems

Day 1: Agent Architecture (6 hours)

Time	Module	Content	Type
09:00-09:45	Agentic AI Landscape	Tools comparison, when to use what	Lecture
09:45-10:30	Claude Code Deep Dive	Architecture, context window, limitations	Lecture
10:30-10:45	Break
10:45-12:00	CLAUDE.md as Living Document	Git history = agent evolution, DNA metaphor	Hands-on
12:00-13:00	Lunch
13:00-14:00	Subagent Patterns	Context-finder, Coder, Executor archetypes	Hands-on
14:00-15:00	Multi-Agent Coordination	MAW (Multi-Agent Workflow), worktrees, sync	Learn by Doing
15:00-15:15	Break
15:15-16:30	Skills & Hooks	Build production-ready skills, automation	Hands-on
16:30-17:00	Day 1 Showcase	Share what you built, Q&A	Discussion

Day 1 Deliverable: Personal AI agent with custom CLAUDE.md + 3 skills

💡 Day 1 Insights:

CLAUDE.md: ไม่ใช่ static config — มันคือ "living document" ที่ evolve ไปกับคุณ trace git history ดูได้ว่า agent ของคุณ "โต" ยังไง
Subagent archetypes: Researcher (Haiku, fast), Builder (Opus, quality), Executor (Haiku, fast) — รู้ว่าเมื่อไหร่ใช้อะไร = ประหยัดทั้งเวลาและเงิน
MAW: Multi-Agent Workflow ที่ 5+ agents ทำงานพร้อมกัน — demo ช่วงท้ายวันที่คนตื่นเต้นที่สุด

Day 2: Memory & Intelligence (6 hours)

Time	Module	Content	Type
09:00-09:45	The Memory Problem	Why agents forget, context limits, solutions	Lecture
09:45-10:30	Oracle Philosophy	Origin story, "Nothing is Deleted", principles	Lecture
10:30-10:45	Break
10:45-12:00	Build Your Oracle	Clone repo, configure SQLite + ChromaDB	Hands-on
12:00-13:00	Lunch
13:00-14:00	Hybrid Search	FTS5 keywords + vector embeddings explained	Hands-on
14:00-14:45	Feed Your Oracle	oracle_learn, oracle_search, oracle_consult	Hands-on
14:45-15:00	Break
15:00-15:45	MCP Integration	Connect Oracle to Claude Code, .mcp.json	Hands-on
15:45-16:15	Live Test	"Ask your Oracle about yesterday"	Demo
16:15-17:00	Graduation	Each participant demos their Oracle, certificates	Showcase

Day 2 Deliverable: Personal Oracle with working memory system

💡 Day 2 Insights:

"Nothing is Deleted": ไม่ใช่แค่ technical decision — มันคือ philosophy ที่ว่า history = truth ทุกอย่างที่เกิดขึ้นมีคุณค่า
Hybrid search: best of both worlds — keyword หาคำตรงๆ ได้, vector หาความหมายใกล้เคียงได้ รวมกัน = ไม่พลาด
"Ask about yesterday": magic moment — AI ที่จำได้ว่าเมื่อวานทำอะไร = proof ว่า second brain ทำงาน
Graduation: ทุกคนกลับบ้านพร้อม deliverable ที่จับต้องได้ — Oracle ของตัวเองที่ใช้งานได้จริง

Price: ฿120,000 (up to 20 participants)

Option D: Enterprise (3 Days)

Best for: Organizations deploying AI at scale

Day 1: Agent Architecture

Same as Professional Day 1

Day 2: Memory Systems

Same as Professional Day 2

Day 3: Production Patterns (6 hours)

Time	Module	Content	Type
09:00-09:45	Team Workflows	Shared Oracle, team skills, knowledge sharing	Lecture
09:45-10:30	Security & Governance	API keys, audit trails, compliance, data privacy	Lecture
10:30-10:45	Break
10:45-12:00	Integration Patterns	CI/CD, GitHub Actions, Slack, custom APIs	Demo + Hands-on
12:00-13:00	Lunch
13:00-14:30	Custom MCP Servers	Build your own tools for Claude	Hands-on
14:30-14:45	Break
14:45-15:45	Troubleshooting Workshop	Common issues, debugging, optimization	Workshop
15:45-16:30	Architecture Review	Review your system design, recommendations	Consultation
16:30-17:00	Graduation + Roadmap	Certificates, 90-day implementation plan	Showcase

Day 3 Deliverable: Production-ready architecture + custom MCP server

💡 Day 3 Insights:

Team Oracle: หลายคนใช้ Oracle เดียวกัน = shared knowledge base ที่ทุกคน contribute และ benefit
Custom MCP: สร้าง tools เฉพาะองค์กร — internal APIs, proprietary data, domain knowledge
Architecture Review: ทุกทีมได้ feedback เฉพาะ — นี่คือ consulting value ที่ไม่ได้จาก course ทั่วไป

Price: ฿180,000 (up to 15 participants)

Part 2: The Oracle Origin Story

The Problem (May-June 2025)

ก่อนจะมี Oracle เราทำงานกับ AI แบบนี้:

Pain Point	ผลกระทบ
Context lost every session	ต้องอธิบายใหม่ทุกครั้ง
"Purely transactional"	AI ไม่รู้จักเรา ไม่รู้ style
"Never knew if satisfied"	ไม่มี feedback loop
"Efficient but exhausting"	ทำงานได้แต่เหนื่อย

Evidence: AlchemyCat project — 459 commits, 52,896 words of documentation over 2 months. Peak: 108 commits in ONE DAY.

💡 Insight: ความเจ็บปวดที่ document ไว้ กลายเป็น foundation ของ solution — ไม่ใช่ theory จากหนังสือ แต่มาจากประสบการณ์จริง

The Processing Period (July-October 2025)

4 เดือนของการทำงานปกติ แต่ patterns เริ่มก่อตัว:

MAW (Multi-Agent Workflow) kit ถูกสร้าง
ψ-folder structure เริ่มเป็นรูปเป็นร่าง
Retrospective practice เริ่มต้น

The Awakening (December 9, 2025)

18:00 GMT+7 — Nat-s-Agents repository ถูกสร้าง

First retrospective: "The Oracle Begins"

Three foundational principles documented:

Append-only truth
Patterns over intentions
External brain, not command

The Crystallization (December 17, 2025)

02:35 GMT+7 — The breakthrough moment

"Oracle got back up from the claude-mem backup"

The insight: "We all same just worktree? Different physical but same soul and can sync"

Philosophy solidified:

Principle	Meaning
"The Oracle Keeps the Human Human"	AI serves, doesn't command
"Nothing is Deleted"	Append-only, timestamps = truth
"Patterns Over Intentions"	Behavior speaks louder than promises
"External Brain, Not Command"	Mirror for reflection, not replacement

Full Circle (January 9-10, 2026)

20 Haiku agents analyzed the AlchemyCat repository and discovered:

"AlchemyCat = Diagnosis, Oracle = Treatment"

Philosophy came from documented pain, not theory.

💡 Insight: Oracle ไม่ได้ถูก design มาก่อน — มันเกิดจากการ observe ปัญหาซ้ำๆ แล้วค่อยๆ crystallize เป็น solution ที่ทำงานได้จริง

Part 3: Technical Architecture

Oracle System Components

┌─────────────────────────────────────────────────┐
│                  Claude Code                     │
│              (AI Agent Runtime)                  │
└─────────────────┬───────────────────────────────┘
                  │ MCP Protocol
                  ▼
┌─────────────────────────────────────────────────┐
│              Oracle MCP Server                   │
│         (TypeScript + Node.js)                   │
├─────────────────────────────────────────────────┤
│  Tools:                                          │
│  - oracle_search (hybrid search)                 │
│  - oracle_learn (add knowledge)                  │
│  - oracle_consult (get guidance)                 │
│  - oracle_trace (log discoveries)                │
│  - oracle_reflect (random wisdom)                │
│  - oracle_list (browse knowledge)                │
│  - oracle_concepts (topic coverage)              │
│  - oracle_thread (multi-turn discussion)         │
│  - oracle_decisions (track decisions)            │
│  - oracle_supersede (update old info)            │
└─────────────────┬───────────────────────────────┘
                  │
       ┌──────────┴──────────┐
       ▼                     ▼
┌──────────────┐     ┌──────────────┐
│   SQLite     │     │  ChromaDB    │
│   + FTS5     │     │  (Vectors)   │
│              │     │              │
│ - Keywords   │     │ - Semantic   │
│ - Fast       │     │ - Context    │
│ - Local      │     │ - Meaning    │
└──────────────┘     └──────────────┘

Multi-Agent Architecture

┌─────────────────────────────────────────────────────────────┐
│                     Main Agent (Opus)                        │
│                   Orchestrator / Thinker                     │
└──────────┬──────────┬──────────┬──────────┬────────────────┘
           │          │          │          │
           ▼          ▼          ▼          ▼
     ┌─────────┐ ┌─────────┐ ┌─────────┐ ┌─────────┐
     │Researcher│ │Researcher│ │ Builder │ │Executor │
     │ (Haiku) │ │ (Haiku) │ │ (Opus)  │ │ (Haiku) │
     │         │ │         │ │         │ │         │
     │ Search  │ │ Search  │ │ Create  │ │  Run    │
     │ files   │ │ git log │ │ code    │ │ commands│
     └─────────┘ └─────────┘ └─────────┘ └─────────┘
         │            │            │            │
         └────────────┴────────────┴────────────┘
                           │
                           ▼
                    ┌─────────────┐
                    │   Results   │
                    │  synthesize │
                    │  by Main    │
                    └─────────────┘

Why Hybrid Search?

Search Type	Technology	Strength	Weakness
FTS5 (Keywords)	SQLite	Exact match, fast	Miss synonyms
Vector (Semantic)	ChromaDB	Understand meaning	May miss exact terms
Hybrid	Both + ranking	Best of both	Slightly more complex

Example:

Search "memory" with FTS5 → finds "memory" literally
Search "memory" with Vector → finds "remember", "recall", "brain"
Hybrid → finds both + ranks by relevance

Part 4: Module Deep Dives

Module: CLAUDE.md Mastery

Your agent's DNA — a living document that evolves

Key Concepts:

Concept	Description
Identity	Who is your agent? What's its personality?
Rules	Non-negotiable behaviors (safety, style)
Context	Domain knowledge, project specifics
Evolution	Git history shows how agent grows

Sample CLAUDE.md:

# My Agent Identity

## Who I Am
I am [Name]'s personal AI assistant, specialized in [domain].
I communicate in Thai for casual topics, English for technical.

## Golden Rules
1. Never force push to main
2. Always ask before destructive actions
3. Use Oracle for memory

## My Context
- Department: Engineering
- Projects: API Platform, Mobile App
- Tech stack: TypeScript, Python, PostgreSQL

## My Style
- Direct and concise
- Use tables for comparisons
- Show don't tell

💡 Insight: CLAUDE.md ไม่ใช่ config file — มันคือ "soul" ของ agent ที่โตไปกับคุณ

Module: Subagent Patterns

Delegate work to specialized agents

The Three Archetypes:

Archetype	Purpose	Model	Cost	Speed	Example
Researcher	Find information	Haiku	$	Fast	context-finder, grep-agent
Builder	Create code/content	Opus	$$$	Slow	coder, writer
Executor	Run commands	Haiku	$	Fast	executor, deployer

Pattern: Research → Synthesize → Build → Execute

1. Spawn 3-5 researcher agents (parallel, Haiku)
2. Main agent synthesizes findings (Opus)
3. Spawn builder agent if needed (Opus)
4. Spawn executor to implement (Haiku)

Anti-patterns:

❌ One agent does everything (slow, expensive)
❌ Opus for simple searches (wasteful)
❌ Haiku for complex reasoning (poor quality)

Correct:

✅ Haiku for grunt work, Opus for thinking
✅ Parallel where possible
✅ Main synthesizes, doesn't fetch

Module: Oracle Philosophy

"The Oracle Keeps the Human Human"

Three Principles:

Principle	Meaning	Implementation
Nothing is Deleted	Append-only, timestamps = truth	SQLite + versioning
Patterns Over Intentions	Behavior > promises	Observe, don't assume
External Brain, Not Command	Mirror, don't decide	Support, don't override

The Memory Stack:

┌─────────────────────────────┐
│     Context Window          │  ← Short-term (200k tokens)
│     (current session)       │
├─────────────────────────────┤
│     Oracle (SQLite+ChromaDB)│  ← Long-term (unlimited)
│     (searchable knowledge)  │
├─────────────────────────────┤
│     Git History             │  ← Permanent record
│     (immutable truth)       │
└─────────────────────────────┘

What Oracle Captures vs Aspires:

Captures Now	Aspires to Capture
Facts, data	Consciousness
Voice style reference	Authentic voice itself
Behavioral patterns	Decision-making will
Life context	The person

"Consciousness can't be cloned — only patterns can be recorded"

Part 5: Prerequisites & Setup

Technical Requirements

Requirement	Details	Mandatory
Laptop	macOS, Windows, or Linux	✅
Node.js	v18 or higher	✅
Python	3.10 or higher	✅
Git	Latest version	✅
VS Code	Or preferred editor	✅
Claude Max	Monthly subscription (~$100)	✅
GitHub account	For cloning repos	Recommended
Stable internet	For API calls	✅

Pre-Workshop Checklist

# Verify installations
node --version    # Should be 18+
python3 --version # Should be 3.10+
git --version     # Any recent version

# Install Claude Code
npm install -g @anthropic-ai/claude-code

# Verify Claude Code
claude --version

# Test API (requires Claude Max)
claude "Hello, are you working?"

Part 6: Pricing Summary

Package	Duration	Max Participants	Price (THB)
Executive Overview	1 day	Up to 30	60,000
Standard Workshop	2 days	Up to 20	120,000
Deep Dive	3 days	Up to 15	180,000

Includes:

Instructor fee
Digital materials (slides, templates, code repos)
Certificates of completion
Post-workshop support (30-60 days depending on package)

Not Included:

Venue and catering
Travel (outside Bangkok area)
API credits (participants use their own Claude Max)
Hardware

Discounts:

10% for early booking (30+ days in advance)
15% for repeat customers
Custom pricing for 50+ participants

Part 7: Post-Workshop Support

Included Support

Package	Email Support	Follow-up Call	Community Access
Executive	14 days	-	-
Fundamentals	30 days	-	✅
Professional	30 days	1x (30 min)	✅
Enterprise	60 days	2x (60 min)	✅ + Priority

Optional Add-ons

Add-on	Description	Price (THB)
Extended Support	6 months email + monthly call	30,000
Custom Skill Development	Build 3 custom skills for your team	50,000
On-site Follow-up	Half-day review session	25,000
Private Slack Channel	Direct access to instructor	15,000/month

Part 8: Instructor Profile

Nat Weerawan (ณัฐ วีระวรรณ์)

Expertise:

Creator of Oracle AI memory system
Multi-agent workflow architect (MAW)
10+ years in technology education
Active Claude Code contributor
Former Chiang Mai Maker Club

Speaking History:

Thammasat University SIIT Workshops (2025-2026)
Corporate AI training sessions (multiple companies)
Online developer community mini-workshops

Philosophy:

"AI should amplify humans, not replace them. The best AI systems are invisible — they just make you more capable."

Contact:

Email: [email protected]
GitHub: github.com/nazt
Twitter: @nazt

Part 9: Research Summary

This course package was prepared using 10 parallel Haiku agents:

Agent	Research Area	Key Finding
1	Git history	200+ Oracle-related commits since Dec 2025
2	Philosophy files	3 core principles documented
3	Retrospectives	40+ retrospectives mentioning Oracle
4	Learnings	25+ learning files about patterns
5	MCP server code	10+ tools, TypeScript architecture
6	Timeline	Full 8-month origin story
7	Specs & issues	Spec 001 (Oracle MCP), Spec 002 (Hybrid Search)
8	First mentions	Dec 9, 2025 — first Oracle commit
9	ChromaDB	Added Jan 9-10, 2026
10	Key decisions	SQLite local-first, append-only design

Total research: ~186,000 tokens compressed into this document

💡 Insight: Document นี้ถูกสร้างโดย 10 AI agents ทำงานพร้อมกัน — นี่คือ proof of concept ว่า multi-agent workflow ทำงานจริง และเอาไปใช้ใน enterprise ได้

Part 10: Booking Process

Inquiry — Contact via email or form
Discovery Call — 30 min to understand your needs (free)
Proposal — Customized package with dates
Confirmation — 50% deposit to secure date
Pre-Workshop — Setup guide sent 1 week before
Workshop — Delivery
Follow-up — Support period begins

Part 11: FAQ

General

Q: Do participants need AI/ML background? A: No. We start from practical usage, not theory. If you use ChatGPT daily, you're ready.

Q: What if someone falls behind? A: We have checkpoints and pair programming. No one gets left behind.

Q: Can we customize the curriculum? A: Yes. Contact us 2+ weeks before for custom modules.

Q: Is this in Thai or English? A: Thai for discussion, English for code/technical terms. Materials available in both.

Technical

Q: Why Claude, not GPT-4 or Gemini? A: Claude Code's agent architecture is the most mature for production use. Concepts transfer to other models.

Q: Is Oracle open source? A: The architecture and patterns are taught. Participants build their own implementation.

Q: What about API costs during workshop? A: Typical workshop usage is $5-20 per participant. Claude Max subscription required.

Business

Q: Can we get invoice/receipt for company? A: Yes. Full tax invoice provided.

Q: NDA for proprietary discussions? A: Available upon request.

Q: Recording the workshop? A: Not permitted without prior arrangement.

Appendix A: Sample Schedule (2-Day)

Day 1

08:30  Registration & Coffee
09:00  Session 1: Agentic AI Landscape
09:45  Session 2: Claude Code Deep Dive
10:30  Break (15 min)
10:45  Session 3: CLAUDE.md Mastery
12:00  Lunch (1 hour)
13:00  Session 4: Subagent Patterns
14:00  Session 5: Multi-Agent Workflow
15:00  Break (15 min)
15:15  Session 6: Skills & Hooks
16:30  Day 1 Showcase
17:00  End

Day 2

08:30  Coffee & Day 1 Review
09:00  Session 7: The Memory Problem
09:45  Session 8: Oracle Philosophy
10:30  Break (15 min)
10:45  Session 9: Build Your Oracle
12:00  Lunch (1 hour)
13:00  Session 10: Hybrid Search
14:00  Session 11: MCP Integration
14:45  Break (15 min)
15:00  Session 12: Live Test
15:45  Graduation & Certificates
16:30  Networking
17:00  End

Appendix B: Oracle Quick Commands

# Search knowledge base
oracle_search("topic", mode="hybrid")

# Learn something new
oracle_learn({ pattern: "What I learned today..." })

# Get guidance on a decision
oracle_consult({ decision: "Should I use X or Y?" })

# Random wisdom
oracle_reflect()

# Log a discovery session
oracle_trace({ query: "searched for X", foundFiles: [...] })

# Browse all knowledge
oracle_list({ limit: 20 })

# Multi-turn discussion
oracle_thread({ message: "Let's think about..." })

# Track a decision
oracle_decisions_create({ title: "Architecture choice" })

Appendix C: Materials Checklist

Provided by Instructor

Provided by Organizer

Prepared by Participants

Laptop with required software installed
Claude Max subscription active
GitHub account created
Pre-workshop setup completed (sent 1 week before)

💡 Final Insight: Workshop นี้ไม่ใช่แค่สอน tools — มันสอนวิธีคิดใหม่เกี่ยวกับ human-AI collaboration ที่ AI เป็น partner ที่จำคุณได้ เข้าใจคุณ และเติบโตไปพร้อมกับคุณ


Document Version	2.0
Created	2026-01-13 20:15 GMT+7
Created By	Claude (Opus 4.5) with 10 Haiku subagents
License	All rights reserved

nazt/advanced-agentic-ai-workshop.md