🔍 The Core — 90% of execs choose wrong

🛠 Workflow — Full cycle 5 steps

Discovery (Code) → Prototype (Code) → Validation (human) → Automation (Code→Cowork) → Operations (Cowork) → Loop.

Cowork runs the code that Code built. Code analyzes the systems that Cowork operates.

💼 Real-world — JANDA 6-month cost

🎯 Claude Code · Cowork advantages and fusion — JANDA's combination pattern

In Chapter 4, I covered 3 interfaces. Among them, the combination of Code and Cowork is JANDA's core. Each excels at different things, and Limitlessman fused the two to build a real system.

💻 Claude Code advantages — AI colleague

• Fast prototyping — Create mockups, HTML, code in 5 minutes and test immediately. The CEO watches alongside and adjusts.
• Human intervention possible — Pause and verify at every step. Maximum safety and accuracy.
• Direct MCP connection — From the IDE, read and write directly to Notion, GitHub, Figma, Supabase.
• Strong at creative work — New mockups, new planning, new design—Code is unbeatable.
• Retrospective and analysis — Analyze operation logs, improve next cycle, organize memory.

🤖 Claude Cowork advantages — AI employee

• 24/7 autonomous — Works without humans via cron and webhooks. Operates even in the middle of the night.
• Infinite scaling — Scale from 1 PM2 worker to 24 simultaneous workers.
• Event response — Responds instantly to Slack, Discord, email, webhook triggers.
• Uniform costs — Stable operational expenses after setup. JANDA = $30/day.
• Structured repetition — Content publishing, CS responses, KPI tracking—Cowork is best-in-class.

🎭 The real difference between two interfaces

🔄 JANDA's fusion — 5-stage full cycle

Using them separately gives 50%. Fusing them gives 100%. Core rule — Cowork operates the code Code builds. Code analyzes and improves the system Cowork operates.

[Stage 1 · Discovery] Claude Code (colleague)
   CEO and user interviews, data exploration, problem definition
       ↓
[Stage 2 · Prototype] Claude Code
   Quick mockup in HTML, Python, Figma (5 min~30 min)
       ↓
[Stage 3 · Validation] Claude Code + CEO
   Review results and iterate — 1~3 rounds
       ↓
[Stage 4 · Automation] Code → Cowork transition
   Convert validated pattern to Agent SDK and PM2 workers
   Deploy (Render, local, VPS)
       ↓
[Stage 5 · Operations] Claude Cowork (employee)
   24/7 autonomous operation → Send results to Slack, email, Discord
       ↓
[Loop ← 4] Analyze operation logs with Code → improve → redeploy Cowork

💎 Operations measured — 1-week content cycle

Code alone = can publish 5 posts weekly but need people daily. Cowork alone = auto daily but no new drafts. Fusion = 5 posts + 7-day auto-post + autonomous CS + log learning.

🎯 Decision Guide — Which to use

Limitlessman does — code with Code, operate with Cowork, analyze with Code, redistribute with Cowork. This loop is the heart of JANDA.

📂 Source: Operational logs [Log] 20260523_code_cowork_optimal_combo_plan.md, [Log] 20260523_claudecowork_vs_claudecode_colleague_vs_employee.md.

💡 "Same Claude, different interface. Wrong choice = 3x cost."

🔍 Essence — API vs MCP, two standards that look same but differ

🛠 Workflow

// claude_desktop_config.json
{
  "mcpServers": {
    "notion": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-notion"],
      "env": {"NOTION_API_KEY": "secret_xxx"}
    }
  }
}

🎨 What Limitlessman did with Figma MCP — no design outsourcing

I have no designer. Yet I had to build AITF API 19-product sales page (1440×11,686px, 12 sections). Figma MCP was the answer.

Flow: CEO request → Claude Code calls Figma MCP's generate_figma_design → 12 sections auto-generated → HTML conversion → posted to Discord #design → CEO edits directly in Figma. Planning 1 hour + Figma MCP 30 minutes + review 30 minutes = 2 hours complete. Designer outsourcing: 2 million won + 2 weeks.

However — there are 5 pitfalls. (1) Image import unavailable, (2) partial generation on atomic task failure, (3) Pretendard font not default, (4) design token auto-mapping incomplete, (5) awkward Korean line breaks. All workaroundable, but — "Figma MCP is not a silver bullet" — you need to know going in.

So JANDA's rule: fast text pages use Notion, polished design proposals use Figma MCP, B2B print materials use Figma MCP + PDF.

📂 Source: Operations records [기록] 20260522_피그마_판매페이지_AI_디자인_시스템.md, [기록] 20260522_피그마_MCP_장단점_분석.md. External reference: modelcontextprotocol.io (Anthropic official, 2024-11).

📓 Notion + AI Agent Full Cycle — Check · Record · Search · Report

Throughout the 5 volumes, I often see "Sihwang auto-records in Notion." Here's the full cycle of how that works — built by me over 6 months.

🔌 API + MCP Used Simultaneously

I handle Notion in two ways at once. Automation via API, conversational search via MCP.

🗄 DB Agent Collects and Saves to Tables

JANDA's Notion has 5 category databases. Each DB — dedicated agent auto-collects, classifies, and saves daily.

Daily auto-collection flow:

[00:30 AM] overnight worker
   ↓ scans 22 markdown records → auto-converts new records to Notion
[06:00 AM] knowledge-ingestion
   ↓ collects external knowledge → chunks to Notion RAG DB
[08:00 AM] silent-trend → 5 new leads to sales DB
[10:00 AM] silent-security → security check → decision log
[12:00 PM] silent-insight → CX patterns → decision log
[17:00 PM] silent-data → KPI tracking → decision log
[19:00 PM] lead-collector → GitHub · Wanted · TheVC new leads
[Every Monday 09:00 AM] weekly-pmo-report → comprehensive report

📥 Markdown → Notion Auto-Conversion (Real Flow of Operations 22 Records)

Limitlessman leaves a [Record] *.md file every day. People don't move it to Notion manually. An AI agent auto-converts it.

def md_to_notion(md_text, parent_id):
    blocks = []
    for line in md_text.split("\\n"):
        if line.startswith("# "):    blocks.append({"type":"heading_1", ...})
        elif line.startswith("## "): blocks.append({"type":"heading_2", ...})
        elif line.startswith("- ["): blocks.append({"type":"to_do", "checked":..., ...})
        elif line.strip():           blocks.append({"type":"paragraph", ...})
    requests.post("https://api.notion.com/v1/pages",
        headers={"Authorization":f"Bearer {TOKEN}", "Notion-Version":"2022-06-28"},
        json={"parent":{"page_id":parent_id}, "children":blocks})

Over 6 months, 22 markdown records → 22 Notion pages auto-converted. Zero JANDA hands involved.

🎨 Full Workflow — Figma + Notion + Email Integration (Proposal Writing·Sending·Approval)

This is the full cycle of Limitlessman sending a proposal to a new client. The CEO's hands touch it just once — a single PIN entry.

[Step 1] CEO types briefly in Discord #sales channel
   "Make a proposal for an 80-person shoe OEM company"
       ↓
[Step 2] PMO bot delegates to Sihwang
   Sihwang calls mcp_notion_search from 18 tools
   → Search Notion for 5 past decision logs in same industry (manufacturing)
   → Learn proposal pattern
       ↓
[Step 3] Figma MCP — invoke design agent
   generate_figma_design (auto 12 sections)
   → Hero · Problem Definition · Solution · 6 Domains · Case Study · Pricing · Next Steps
   → Generate 1440×11,686px proposal in under 30 minutes
       ↓
[Step 4] Notion API — auto-create child page in decision log
   "Shoe OEM Proposal v1 — 2026-05-27"
   → Auto-embed Figma share link + key content
   → Grade L = L3 (external impact)
       ↓
[Step 5] Auto-generate email body (Claude Sonnet)
   "Hello, CEO OO. This is Limitlessman at JANDA.
    Over the past 6 months, we've documented shoe OEM cases..."
   → Generate body only with preview_mail tool (no sending)
   → Post to Discord #preview channel
       ↓
[Step 6] CEO review + revision comments
   "Remove one line on pricing, add case study to first line"
   → Sihwang edits body → 2nd preview
   → 2–3 rounds possible
       ↓
[Step 7] CEO final PIN entry
   send_mail_with_pin tool
   PIN = "JANDA does this" (company name exact match)
   → Typo = no send. Exact match = SMTP send
       ↓
[Step 8] Auto-record after sending
   · Notion decision log: "Sent, 2026-05-27 14:32"
   · Discord #cs: "📧 OO Company proposal sent"
   · Sales lead DB: status auto-updated to "proposal stage"
   · 5-minute watchdog: alert follow-up if no response

From Step 1 (one-line Discord input) through Step 8 (auto-record) — CEO time total 7 minutes (Step 1 30 seconds + Step 6 5 minutes + Step 7 10 seconds + review 1 minute). The same task split across designer + copywriter + sales manager (3 people) would take 2–3 days average.

This is the real full cycle of Notion + AI agent. When each tool does its job — the whole cycle frees the CEO's hands.

📁 Google Drive → Meeting Notes → Notion Auto-Routing

Limitlessman just drops meeting recordings into a Google Drive folder. Then auto-routing takes over.

[Step 1] CEO uploads wav/mp3 to Google Drive folder on mobile after meeting ends
   ↓ Drive API webhook (or 1-minute polling)
[Step 2] meeting_processor worker detects
   → Drive API: fetch new file metadata + download URL
[Step 3] Whisper API STT ($0.36/hour)
   → Full transcript + speaker separation
[Step 4] Claude Haiku 4.5 — classify into 5 categories
   → Decisions · To-dos · Issues · Opportunities · Risks
[Step 5] Notion API → auto-create child page in meeting notes DB
   → Title · Attendees · 5 categories · full transcript · original wav link
[Step 6] Extract decisions & to-dos → auto-register in decision log DB + task DB
[Step 7] Discord #ceo-briefing alert ("1 meeting auto-summarized")

1-hour meeting = STT $0.36 + Haiku classification $0.001 + Notion API free = $0.37. Same task by secretary takes 1 hour. Janda meeting forget rate 87% → 0%.

📂 Source: Operations log [Record] 20260521_NotionIntegration_SelfGeneratingEnhancedRAG.md, [Record] 20260522_NotionIntegrationDB_OpenAIFullSet_MeetingAutomation.md, [Record] 20260523_Market_Chat_OperationsRoom_PermissionMatrix_v3.md. Actual operations code: meeting_processor.mjs, cmd_processor.mjs, weekly_pmo_report.mjs. Notion parent page ID: 368a9728-88c0-811c-9955-e524eabd1583. Google Drive MCP·API: developers.google.com/drive.

💡 "MCP is the USB-C of the AI era. If you don't know it, you'll be behind in 5 years."

🔍 Essence — Model Routing

💼 Practice — Model Router 50 Lines

def route(task: str) -> str:
    # Step 1: Haiku classification
    cat = haiku_classify(task)
    if cat in ["simple", "format", "extract"]:
        return haiku_respond(task)
    elif cat in ["analyze", "synthesize"]:
        return sonnet_respond(task)
    else:  # "decide", "complex"
        return opus_respond(task)

📊 Limitlessman's Content Reuse — 4 Ebooks → 64 PPT Slides Automated

Limitlessman published 4-volume ebooks, then needed to convert each into PPT presentations. For lectures, presentations, and SNS slide decks. Direct approach would cost ₩500K–1M per volume × 4 = ₩2–4M, 1–2 weeks. The Janda way was different.

Flow: Book .md source → Haiku auto-summarizes 4 slides per chapter → python-pptx auto-generates slides → matplotlib auto-creates 9 diagrams → HTML viewer integrates. 4 book PPTs (64 total slides) + 9 diagrams + viewer = ~1 hour, Claude API $3.

One resource became five. 1 ebook = 1 PPT = 18 videos = 18 blog posts = 54 SNS cards. A year's worth of content extracted from a single published book. This is one reason Limitlessman achieves 30-person impact as a solo operator.

📂 Source: Operations log [Plan] 20260526_PPT_Diagrams_20Each_PerVolume.md, slides.json 64-slide extraction data (2026-05-26). Libraries: python-pptx, matplotlib.

💡 "1 Sonnet = 3 Haiku. Opus only for difficult decisions."

At first, it was great.

A month after Limitlessman discovered Claude. In the morning: "Summarize today's trends for me." At lunch: "Categorize this meeting transcript." In the evening: "Five concept sketches for tomorrow's content." Before bed: "Draft replies to my emails." I was calling Claude around thirty times a day.

Time flew. Output multiplied. In just one month, I saw the effect of having one more full-time employee.

But then—in the second month, something felt off.

Limitlessman took a vacation. Four days. Zero AI calls.
Limitlessman fell asleep. Eight hours. AI did nothing.
Limitlessman was in a meeting. Four hours. AI was idle.

AI was waiting for the boss.

Of the thirty daily calls—when I looked closely, twenty-seven were repetitive tasks. Trend summaries, transcript categorization, email drafts, KPI checks. Same time, same pattern every day. These were jobs the AI could do on its own without the boss having to ask.

Andrew Ng said this somewhere: "Manual AI invocation is the bottleneck of productivity in the AI era." The GitHub Octoverse 2024 report confirms it—developers average 80 AI calls per day, and 70 percent of them could be automated.

Limitlessman looked at the clock that day. I was spending two hours a day on these calls. Sixty hours a month. Half a full-time employee's worth.

AI delivered the effect of one more employee, but the boss's time doubled. Strange, wasn't it? Using AI cost me more time.

"When AI waits for the boss—the boss becomes the bottleneck."

So the next domain is automation. PM2 · webhook · silent_activator — tools that make AI work even while the boss sleeps.

🔍 The Essence — "Stops When You Close the Laptop" Is Not Automation

The definition of automation is — it works even while people sleep. 24/7 is only possible when separated from people.

There's a psychological concept called Always-On Bias — the boss's compulsion to manage everything personally. PM2 breaks this compulsion.

🛠 Workflow — PM2's 4 Values

• Auto-resurrect when dead (automatic restart on process.exit(1))
• Auto-start on boot (automatic after system reboot)
• Memory protection (max_memory_restart)
• Log accumulation (~/.pm2/logs/)

pm2 start ecosystem.config.cjs
pm2 save              # save current state
pm2 startup           # register auto-start on boot
pm2 list              # list running workers

🛠 The 24 Workers That Limitlessman Actually Automated — Full Specifications

These are the workers that Limitlessman built step by step over 6 months and still runs 24/7 today. Orchestration = a system where 24 workers are connected by time and events.

When these 24 workers run at different times and different events — while Limitlessman operates, I work directly for only 4 hours and 30 minutes. The remaining 19 hours and 30 minutes, the system works. This is "the actual measurement of orchestration."

📂 Source: Operations log [Record] 20260522_Market_Autonomous_Operations_System_Build.md, [Record] 20260525_Security_System_Build_17silent_Diagnosis_3bot_Archive.md, [Record] 20260526_Market_Monthly_Patcher_Acceleration_70Goal.md. Appendix C has full ecosystem.config.cjs code.

💡 "If it stops when the laptop shuts down, it's not automation."

🔍 Essence — Polling 1,440 times vs Webhook 1 time

💼 Practice — 5 Minute Setup

@app.post("/webhook/payment")
async def payment(req: Request):
    data = await req.json()
    # Register in Notion + Discord notification
    notion_add(data)
    discord_post(f"💰 {data['customer']} {data['amount']:,} KRW")
    return {"status": "ok"}

💡 "Polling is control obsession, Webhook is system trust."

🔍 Essence — "AI Receives Work" → "AI Creates Work"

The boss throws work and AI activates? Then the boss becomes a single point of failure.

JANDA's 5 silent workers auto-trigger 5 tasks daily. CEO doesn't command them.

🛠 JANDA 5 Silent Workers

💡 "The first step to making AI not a tool, but an employee."

It was 2 AM on May 20, 2026.

That night, Limitlessman deployed ten new modules to Discord all at once. Tool execution engine. Memory system. Evaluation tournament. Self-evolution loop. Organization hierarchy. Project engine. RAG search included. A seven-layer full stack grew at 1 AM. Commands multiplied to twenty-two. The module exceeded a thousand lines.

It was late, but I wanted to run it just once. I opened Discord and typed !project list.

Crash.

I missed one import line. I forgot one function while renaming. Fixed it. Tried !project create again — type error. Didn't reflect the signature change. Fixed it again. Tried to call the Python bot — no response. I discovered that night that spawn() was ignoring the timeout option. And finally — claude-opus-4-7 was hardcoded inside the RAG engine. A policy violation JANDA had set.

Four bugs exploded at the same time. From 2 AM to 4 AM was debugging. The CEO who built ten modules — was juggling four of them simultaneously.

That dawn, I realized something.

Build ten automations — and those ten automations devour the CEO.
Modules increase, but CEO time doesn't decrease.
It actually increases. Monitoring, management, debugging all fall on the CEO's shoulders.

At first, five workers. Five minutes to check and done. But when it becomes twenty-four workers — checking alone takes an hour and a half. It grows non-linearly. The Stripe Engineering Blog says the same thing. "Run a hundred workers without self-healing, and you need one SRE full-time."

Linear's founder takes it a step further. "AI agents fix agents. That's the secret to a million dollars ARR per employee."

At 4 AM, Limitlessman wrote one line in Notion.

"We need automation of automation."

That's an AI agent.

Agents diagnose themselves. Agents fix their own code. If an agent dies at 2 AM — another agent brings it back to life. The CEO sleeps.

The phase where a CEO builds automation — that's Domain 3.
The phase where automation fixes automation — that's Domain 4.

📂 Source: JANDA Development Log 07_ebook_publishing/development_log.md 2026-05-20 late night entry. Discord gateway 22 commands · 10 new modules · 4 bugs discovered simultaneously.

🔍 The Essence

LLM's first attempt scores 60. People don't know it. Make it critique itself, and it finds the flaws.

"Don't finish in one shot. Look at it again." — LLM version of the universal principle in Korean bestsellers.

🛠 6-Stage Cycle

1. generate (Sonnet first attempt)
2. critique (Haiku self-critique — severity)
3. regenerate (if severity=high, regenerate)
4. validate (subprocess validation)
5. apply (commit) or skip
6. reflexion_memory append (accumulated learning)

def reflexion_loop(task):
    answer = generate(task)            # Sonnet
    critic = critique(task, answer)    # Haiku
    if critic.severity == "high":
        answer = regenerate(task, answer, critic.defects)
    log_to_memory(critic)
    return answer

💡 "32-point gap = 5x business impact. Reflexion = treating AI as a learner."

🔍 The Essence — AI Pays Off Code Evolution Debt

Code without evolution = debt. When the CEO pays it, the CEO regresses to coder. When AI pays it, the CEO evolves to decision-maker.

🛠 8-Stage Cycle

Diagnose → generate(Sonnet) → self-critique(Haiku) → regenerate → subprocess validation → confidence scoring → threshold branch → git commit → memory append

💡 "Let AI pay off code evolution debt."

🔍 The Core — Why Klarna Really Failed

Surface: 700 AIs collapsed. Core truth: No permission matrix existed.

🛠 5-Level Pyramid

💡 "Too little authority, I'm powerless. Too much, I cause accidents. 5 levels is the golden ratio."

It was the day after Limitlessman created ten agents.

The marketing agent was creating content. The CS agent was handling customer support. The sales agent was classifying leads. The data agent was tracking KPIs. The self-coding agent was attempting patches.

All five were doing their jobs well.

But — a week passed.

Limitlessman counted the collaboration messages. Zero. The five agents never spoke to each other once. Each did only their own work.

Marketing created content — but CS didn't know about it.
CS handled customer complaints — but sales didn't know about it.
Sales discovered new leads — but data didn't know about it.
Data tracked KPI anomalies — but self-coding didn't know about it.

This is a scene I saw 30 years ago in a people company. Departmental silos. The marketing team in their silo, the sales team in theirs, the CS team in theirs. Each doing well — but not as a company, rather like five slaves working separately.

The Microsoft Work Trend Index 2025 report says exactly this — "80 percent of AI agents are trapped in silos with zero ROI." The OpenAI Agentic Workflows document reaches the same conclusion — "Multi-agent communication creates 70 percent of agent ROI."

The next week, Limitlessman turned on Discord. He created 30 channels, attached 11 bots, and issued 30 webhooks. A company where five agents hadn't spoken to each other for a week — on the first day of Discord adoption, exchanged 200 collaboration messages.

Just as people companies have KakaoTalk and Slack — AI companies have Discord.

"Building 30 agents is 50 percent. When 30 talk to each other, that's 100 percent."

That's why the next domain is Discord. The infrastructure that breaks agent silos.

🔍 Essence — AI Agents' KakaoTalk

📡 The Bidirectional Channels Limitlessman Actually Runs — Discord + Telegram Separated

Limitlessman operates these two channels with role separation running simultaneously. "AI collaboration" and "CEO command line" should never be on the same channel — that's a rule I learned over 6 months.

JANDA's flow of use is bidirectional.

• CEO → Market Briefing: From Telegram mobile, /market Tell me today's priorities → Market automatically generates brief with 18 tools → Telegram answer in 12 seconds.
• Market → 30 Agents: In Discord #PMO channel, PMO bot calls delegate_agent("MKT") → MKT bot posts results in Discord #mkt channel → blackboard accumulates automatically.
• 30 Agents → CEO: Self-recovery and self-utterance results auto-post in Discord #ceo-briefing channel → CEO receives summary alerts via Telegram (silent_writer organizes every Monday 09:00).

This is bidirectional orchestration. Using only one channel creates silos or noise — separating two channels is JANDA's 6-month conclusion.

📂 Source: Operations Log [Log] 20260520_AGI_Organization_Restructure_Recovery.md, [Log] 20260523_Market_Chat_Operations_Authority_Matrix_v3.md. Actual code: discord_agent_gateway.mjs, telegram_agent_bot.py.

💼 Hands-On — 5-Minute Setup

1. Create 1 Discord server
2. Issue bot token (discord.com/developers)
3. Right-click channel → Webhook → Copy URL
4. Send 1 message via curl

💡 "Discord is the operations channel of the AI era."

🔍 Essence — Sihwang Operations Know-How #2

"Not 30 people, but 3 hubs (PMO·DATA·OPS) are the key."

— One person weathered a 6-month crisis together, and that's why they remained a colleague, not an employee. The starting point of the 30 agent organization chart is not a chart for cutting people, but a chart that maps out spaces for both humans and AI together.

🛠 3-Layer Org Chart

• CEO Limitlessman — Decision-maker
• Sihwang (PMO) — Claude Haiku 4.5 + 18 tools
• 5 PMs — OPS, DEV, MKT, AI, DATA
• 24 Specialists — Self-activate via silent_activator

🤝 4 Collaboration Patterns

1. Vertical Delegation — PMO → Team Lead → Specialist
2. Horizontal Sharing — Blackboard append-only
3. Self-Activation — silent_activator 5 times daily
4. Self-Recovery — health-monitor → REVIEWER → AUTO

📅 Limitlessman's Day — Timeline Log

I've covered tools, systems, and architecture enough. Finally — how I actually work on one real day. No fiction.

00:30 — overnight auto-start (28 rounds self-learning)
02:30 — patcher-force: Sihwang diagnoses own code (confidence 78%)
06:00 — knowledge-ingestion: 12 external knowledge items → Notion RAG
08:00 — silent-trend: "Claude 4.8 launch planned" auto-discovered → Discord

08:15 — CEO checks Sihwang on mobile
   "Today's priority?"
   ↓ Sihwang Haiku 4.5 runs read_state tool (8 seconds)
   ↓ "AITF API response delay 6552ms — ROI 30, 0.5h. Priority 1."

09:30 — Claude Opus (IDE) ↔ Sihwang discussion
   Opus: "Patcher threshold 80%, lower to 75%?"
   Sihwang: "75 is risky. 3 amem failure cases. 78% + backup reinforcement recommended."
   Opus: "Agreed. Making PR."
   CEO: "OK" (1-click)

11:30 — New customer proposal auto-generated
   CEO: "Make shoe OEM proposal"
   ↓ PMO → DESIGN → Figma MCP (12 sections in 30 min)
   ↓ Notion auto-sync + Discord alert

12:00 — silent-insight: blackboard synthesis → insight
14:00 — preview_mail: CEO review → "Skip price, fix case intro line"
14:32 — send_mail_with_pin: CEO PIN "Janda proceeding" → SMTP dispatch
17:00 — silent-data: KPI auto-tracking
19:00 — lead-collector: 47 new leads
21:00 — CEO wrap-up: 5-min Discord scan
22:00 — health-monitor: 24 workers OK
Midnight — overnight restart. I'm sleeping but the system keeps working.

CEO hands-on time total = roughly 22 minutes. The remaining 23 hours 38 minutes — 24 agents and 24 workers do their jobs.

This is a real day for a solo founder CEO + 30 AI. Not tools — it's a flow of collaboration. Claude Opus deliberates like a colleague, Sihwang Haiku operates like an employee, and 30 agents work asynchronously. I just approve. 22 minutes.

📂 Source: Janda operations log 2026-05-27. All actual times, decisions, and utterances logged in _sihwang_chat_log.jsonl, _blackboard_log.jsonl, _studio_commands.jsonl.

💡 "30 person-years = 6 AI months. 5 PMs fit the founder's brain limit."

🔍 Essence — Survivorship Bias

Discord decisions are unfindable after one week. This is why the CEO repeats the same mistakes.

"Failure rounds must be recorded so AI grows." — Shihwang's Insight #5.

🛠 cmd-processor Flow

Discord decision → 5-second polling → Notion child page auto-create → Action Items to-do block → 5-minute watchdog → No response = CEO alert

💡 "Automatically bridges volatility and permanence."

Essence

30 agents collaborate. 200 work messages a day. But who sets priorities? The CEO every time = CEO work all over again. 30 instruments without a conductor = noise. Orchestra is the answer.

Actual Facts

• JANDA Measurement: Before Shihwang launch, 200 utterances per day → CEO ignored 80% → agent utilization 50%. After Shihwang launch, CEO responds to only 3–5 items per day → utilization 92%.
• Anthropic Agent SDK: "Without orchestrator, multi-agent devolves into noise."
• LangGraph (2024): "DAG orchestration = SOTA for >5 agents."
• Microsoft Agent Framework: "Orchestrator is the heart of enterprise AI."

"30 instruments become music. Conductor = Sihwang. Next domain = Orchestra."

🎯 5 Principles of Task Management — Foundation of the Tool System

To design 18 tools well — first, I need the "5 Principles of Task Management." These 5 principles are implemented as the 18-tool system.

These 5 principles are not JANDA's discovery alone. 7 global authorities say the same thing.

5 principles + 7 authorities + operational field data — this is the design foundation of the 18 tools. "AI era = mastering task management" is not something I alone claim — it's a fact the world agrees on.

📂 Sources: Karpathy Twitter 2023, Mollick 「Co-Intelligence」 2024 (Penguin), McKinsey「The Economic Potential of Generative AI」 2024, Microsoft Work Trend Index 2025.

🔍 The Essence — Choice Paralysis Limit

30 tools → LLM paralysis. 5 tools → powerless. 18 is the golden ratio.

🛠 4 Categories 18 Tools

💡 "18 is the golden ratio. Tool names reflect purpose, not function."

🔍 Essence

"70% of AI errors stem from operator mistakes. Without a permission matrix, it's only a matter of time."

🛠 L3 Mail Gate — 3 Steps

1. preview_mail — Generate body + Discord preview
2. CEO PIN entry — Exact company name match ("JANDA agrees")
3. send_mail_with_pin — SMTP send + log

🌐 Four Cases From Other Companies — "What About Mine?"

The skeptic asks: "JANDA can do it. But my company?" I answer with 4 cases. Find the one closest to you.

🎬 Meta proof — How this book was made

Let me be clear first — this book was not written by AI.

Limitlessman planned it directly, and I organized 6 months of actual operational data myself. However — I used Siyoung (Claude Haiku 4.5 operator AI) in the review process. Like an external editor reviewing one chapter of the book, Siyoung spotted weaknesses from an operator's perspective.

There's a reason why this book was made that way. The 4-book series — the production method evolved one step with each volume. Book 5 is the result of that evolution.

Book 2 taught tools, 3 was agent design, 4 covered 3-axis operations. Book 5 — is the stage of integrating all of that into my own operating system to make the book. That's why Siyoung's review came naturally. If I don't use the system the book teaches while writing the book itself — the book becomes a lie.

Why disclose this in the book? — because book 5's core message is "making AI work for you". If I don't use that system while writing the book, the book becomes a lie. Dogfooding — using your own tool yourself — is the proof.

What follows is the actual record of that review process:

💡 "AI mistakes are operator mistakes. JANDA 6 months zero incidents = proof of the gate."

🛠 Company Size Matrix

🏢 10 Real Companies Deploying AI Agents & Orchestration (2024–2026)

"Do AI agents actually get used in real companies?" — The answer is yes. Global enterprises have been deploying at scale since 2024. While 30 agents is a solo founder case, below are examples from companies with tens of thousands to hundreds of thousands of employees.

💡 10 Enterprise Cases → 5 Key Messages

Looking at the 10 companies above — 2024~2026 was the inflection point for AI agent adoption. The path Klarna took by moving too fast and rehiring people, the path Salesforce took by declaring with Agentforce that "we sell agents, not software," the path JP Morgan took by saving 360,000 hours — they all follow the same flow.

30 Agents is a solo founder case study. The 10 companies above are cases with tens of thousands to hundreds of thousands of employees. Scale differs, but structure is the same. Domain clarity → Data ingestion → AI utilization → Automation → Agents → Discord (or Slack·Teams) → Orchestration.

The 6 domains in Book 5 apply equally to solo founders and Fortune 500 companies. The difference is not tools, but domain clarity and the detail of the authority matrix.

📂 Sources: Salesforce Agentforce 2.0 Official Announcement (2024-10) · CX Today "Klarna Redeploys Staff" (2025) · JP Morgan COIN Official (2017~) · Goldman Sachs AI Assistant Announcement (2024-06) · Citi AI Stylus Announcement (2024-09) · Microsoft Build 2026 Keynote · Anthropic Customer Stories (Moody's·Shopify) · LinkedIn Engineering Blog (2024).

💡 "Framework is a tool. Domain determines the tool."

🏗 9-Step Enterprise Adoption Roadmap — What to Do Before Orchestration

"Can we just build 30 agents and be done?" — the most common question I get. The answer is "No, there are 8 steps before that". Orchestration is the second-to-last of 9 steps. If you skip steps 1~7 and start from step 8 — you'll follow the same path Klarna took, laying off 70,000 people and then rehiring.

Below is the 9-step sequence that doesn't fail, synthesized from 6 months + 10 Fortune 500 companies + Anthropic·McKinsey·Gartner research.

📋 Step-by-Step JANDA Mapping + External Validation

The same 9 steps — triangulated across JANDA lectures, Fortune 500 adoption cases, and academic/vendor research.

⚠️ Skip a Step — 4 Common Failure Patterns

📂 Sources: McKinsey "The State of AI 2024" · Gartner "AI Strategy Trends 2024" · Anthropic MCP Announcement (2024-11-25) · Anthropic "Building Effective Agents" (2024-12) · OpenAI "A Practical Guide to Building Agents" (2025) · LangChain LangGraph Docs (2024) · Microsoft Agent Framework v1.0 Release Notes (2025) · CrewAI Documentation · Air Canada v Moffatt (BC Civil Resolution Tribunal, 2024-02)

🎨 One Picture — 6 Domain × 9 Stage Integrated Matrix

💡 The 9 stages are time sequence (data to permissions). The 6 domains distinguish roles (D1 data through D6 orchestration). The two axes are not the same "9" — when you map the 9-stage time axis onto 6 roles — the following matrix emerges.

The entire book appears in one chart. Horizontal: 9 stages (data → permissions) × Vertical: 6 domains (data → orchestration). Each cell shows which chapter in this book fills that coordinate.

All 18 lectures are in one diagram. I can know exactly which domain and which stage my company is at — more precisely than a persona roadmap. Cafe = D1 ①②, Medical clinic = D1·D2 + ⑨, Shopping mall = D2·D3·D6, JANDA = 6 domains 9 stages full stack.

📂 Visual pattern: James Clear "Atomic Habits 4 Laws × 4 Stages" + James Clear "Atomic Habits 4 Laws × 4 Stages matrix" tone reference.

🩸 One Operational Failure — "I was collecting data, but for 6 months nothing was actually being collected"

return {"total_collected": 0, "agi_learning": "stub"}

📂 Source: 02_AI_Agent_Lab/agent_development/knowledge_ingestion_engine.py stub self-discovery (2026-05-23) · CLAUDE.md ⚠️ flag ("current STUB ... situation actual learning = world_monitor + briefing loop + 30 Discord agents") · [Log] 20260523_*.md series.

💡 "Orchestration is step 8 of 9. Not the starting point — the 6-month finish line."

I made it here. I read all 7 prologue boxes + 18 chapters. 80% of readers close the book at the 1/3 mark. You are not that 80%. This is what sets you apart one year from now.

🌟 My 4 Actions as Author

"To be JANDA—an AI agent company that uses AI better than anyone else, something everyone can use.
I'm working toward that goal. Execute·study·apply·meet."

— The fact that you are asking these 4 things right now means you are already an executor. After you close this book — I'll see you at AI House in one year.

🔥 In the Age Where Everyone Builds — The New Divide

— After closing this book, ask yourself: Do I intend to build an "environment" to spend 100 million tokens by this time next year?
Only those who ask about "environment," not "volume," reach 100-agent orchestration.

💎 One-line analogy: "The cumulative curve of 100 million tokens with environment setup goes further than 10 billion spent recklessly without environment."

※ The above 10 billion/100 million comparison is a metaphorical expression to explain the asymmetry of system accumulation. It does not guarantee actual ROI for individual users.

"This book is not the end. It is the beginning.
A year from now, I'll meet you at AI House."

— Limitlessman

🛠 The Workflow I Actually Built — 6 Domain Detailed Specification

Volume 5 is not abstract. It is the actual workflow I built step by step over 6 months. Below is — what I did in each domain, in what order, and with what tools — full specification.

📌 Core Deliverables — What Limitlessman Has After 6 Months

📂 Sources: 22 operational logs (2026-05-20~27) + development-record.md (31KB) + appendices C·D·E·F full spec + Sihwang 18 tool schemas.

This table is JANDA Book 5's real outcome. Not abstract — from 1 Notion parent page in week one to a 30-agent · 24-worker · 19-product SaaS operator after 6 months. Draw the same table for your company. One year from now, you are Limitlessman inside that table.

🎯 The One Diagram: CEO Agent Connecting All 6 Domains

Book 5 finale — when a CEO (or CEO agent Sihwang) connects 6 domains at once, how does the company operate? One page of measured operational flow.

                    [CEO (or CEO Agent — Sihwang)]
                              │
        ┌─────────┬───────────┼───────────┬─────────┐
        ▼         ▼           ▼           ▼         ▼
    [Notion DB] [Design]   [Discord]   [Email]  [Sales·
    Agent      Agent      Orchestra    Agent     Payment]
                                                 Agent

    │ Notion         │ Figma MCP   │ 30 webhook  │ SMTP        │ jbooking
    │ 450 DB      │ 12 sections     │ 11 bots       │ preview→PIN │ Stripe
    │ 1,421 cases     │ 30min auto   │ 200cases/day    │ L3 gate   │ receipt

        ↓             ↓             ↓             ↓             ↓
    Company OS      Sales Page    AI 30 Collab   Customer Service Revenue Auto
    SSoT search    Proposal·Print   Async flow    L3 secure    Tax auto

        └─────────┴───────────┼───────────┴─────────┘
                              ▼
                  [Integration Result — Auto Daily]
                  · Meeting notes auto to Notion
                  · Sales leads auto-respond (L3 PIN)
                  · Content auto-publish (silent 5)
                  · Self-coding patcher (78% conf)
                  · Discord 200cases/day autonomous collab
                  · Self-recovery 92% (human touch 1-2x/day)

                              ↓
                       [CEO Signs Off Only]
                  L0·L1: Market auto (50cases/hour)
                  L2: 1-click approve (5min)
                  L3: PIN entry (company name exact)
                  L4: CEO direct (ebook·payment·legal)
      

This is the conclusion diagram from all 5 books. 1 CEO + 1 market analyst + 30 agents = 1 company. After 1 year of operation: 100M revenue, 20M operating costs, 0 labor costs, employee satisfaction high—because humans focus on the 20% only humans can do.

📂 Source: 6-month integrated workflow. All components taught step-by-step in Chapters 1~18 + Appendix 9 in the main text. This diagram connects every chapter to the final answer.

💡 The appendix to this book — is fully inside the book. This is not a book whose value is completed only by receiving separate materials. If you can shut this book and run 1 system tonight — that's the appendix, and that's completion.

📘 In-Text Appendix — 3 Core Elements

3 essential items applicable within 5 minutes of closing the book. Fully inlined in the text.

🌳 Series Domain Mapping — What Books 1, 2, 3, 4, and 5 Fill

Book 5 is sufficient as a standalone, but as a series, the weight of the message changes. This is an objective mapping of which domains each book covers.

◎ = Core chapter · ○ = Partially covered · △ = Mentioned in introduction · × = Not included. If you buy only Volume 5, domains 1–6 will have gaps; if you buy only Volume 1, domains 4–7 will be empty. Why the series is a 1-year cumulative manual.

🪞 Before Closing the Book — Narrative·Self-Reflection Tone 3-Question Checklist

🛰 The Book's Content — 3 Places Limitlessman is Running Right Now

Every piece of content in this book is a system I'm actively running at this moment. Here are three places where you can directly verify what's cited in the text. These are verification links, not sales pitches.

📂 22 Operating Records — Primary Sources for Book 5

Book 5 is not a "lecture book" — it's a collection of 22 records that Limitlessman left behind while operating for 6 months. Here's where each chapter comes from.

This is the basis for the book's credibility. Every chapter is written on top of a record where Limitlessman actually did that work before writing about it.

Limitlessman (JANDA) · June 2026
aiaijungle.github.io/vibe-ebook

Lectures 7~18 + Epilogue + 9 Appendices Preview

• Domain 3 Automation (3 lectures) — PM2·Cron, Webhook, silent_activator
• Domain 4 Agents (3 lectures) — Reflexion, self-coding patcher, L0~L4 permissions
• Domain 5 Discord (3 lectures) — gateway, 30-agent organizational chart, decision log archive
• Domain 6 Orchestration (3 lectures) — 18 tools, L0~L4 gates, LangGraph/CrewAI/Microsoft
• Epilogue — author's 4 actions (execution·study·application·connection), AI house in 1 year
• Appendices A~I — Notion DB·30 webhooks·PM2 24 workers·Siyoung 18 tools·autonomy_policy·6-month costs·5 failures·30 copy examples·30 pitfalls·bundle funnel

🌟 Limitlessman's Vision — Author's 4 Actions

"An AI agent company that leverages AI anyone can use to the absolute best.
I'm aiming for the author's way. Execute·Study·Apply·Meet."

• Execute — 5 minutes today. The CEO builds it directly.
• Study — 1 book per week + meetings. Ask if you don't know.
• Apply — 1 case daily. Reading alone leads to forgetting.
• Meet — Monthly meetups + 1 person. You can't go alone.

"Let's meet at AI House in one year."
— Limitlessman