Chapter 2 — Claude Code Deep

🧠

"30 hours autonomous coding. 11,000 lines. Zero intervention." — Claude Sonnet 4.5, Sep 2025

You'll learn

Claude Code architecture: agent loop + tools + MCP
Sub-agent orchestrator-worker pattern (Anthropic Research style)
Case: Jaana Dogan (60 min on problem Google team iterated all of 2024)
Cost economics: Uber CTO burns $500-2K/dev/month
Workflow for devs + global freelance rates

01 What is Claude Code?

Claude Code = Anthropic's terminal-native autonomous coding agent.

Architecture

┌─────────────────────────┐
│   Claude Sonnet 4.6      │
│   (or Opus 4.7)          │
└───────────┬─────────────┘
            │
        Agent loop (ReAct)
            │
   ┌────────┼────────┐
   │        │        │
 Tools    Memory   MCP
(Bash,   (history (3rd-party
 Read,   + plan)   servers)
 Edit,
 Grep)

Pricing (May 2026)

Tier	Cost/month	Use case
Pro	$20	Personal dev, light usage
Max	$100-200	Pro dev, heavy use
API pay-as-you-go	$3/$15 per MTok (Sonnet 4.6)	Custom integration
	$5/$25 per MTok (Opus 4.7)	Heavy / mission-critical

02 Sonnet 4.5 — "30 hours autonomous" moment

Story

Sep 2025, Anthropic released Sonnet 4.5. Biggest demo: 1 Claude Code instance ran:

Metric	Number
Autonomous coding time	30+ hours continuous
Code generated	~11,000 lines
Human intervention	0
Task	Full-stack app from scratch

Why breakthrough?

Before Sonnet 4.5:

GPT-4 / Claude 3.5 autonomous: 1-4 hours before "context drift" or fail
4-8 hours was max human-trust threshold

After Sonnet 4.5:

30+ continuous hours — model maintains context, plans recovery on errors
Anthropic calls this the first true "agentic" model

03 Sonnet 4.6 + Opus 4.7 (May 2026)

Model	Release	Strength	Best for
Sonnet 4.5	Sep 2025	Balance price/perf, 30h autonomous	Daily coding
Sonnet 4.6	Feb 2026	OSWorld 72.5% (≈ human baseline)	Daily coding upgrade
Opus 4.7	Mar 2026	Best reasoning, complex architecture	Critical, mission-critical
Haiku 4.5	Jan 2026	Fast + cheap ($1/$5/MTok)	Sub-agent worker, bulk

04 Case 1 — Jaana Dogan / Google Gemini API team

Jaana Dogan = senior engineer Google Gemini API. She worked on distributed agent orchestration system all of 2024.

Feb 2025 she tried Claude Code.

Item	Number
Problem	Distributed agent orchestration
Google team iterated	All of 2024
Claude Code solo	60 minutes
Output	Working prototype

"Claude Code solved in 60 minutes what my team at Gemini iterated on for a year." — Jaana Dogan

→ Lesson: even senior engineers at top-tier companies get 10x'd by Claude Code. Not replace — augment.

05 Case 2 — Uber CTO burns AI budget in 4 months

Story

Apr 2026 — Uber CTO public report:

Metric	Number
Claude Code adoption	32% → 84% in 5,000-engineer org
Per-engineer API cost	$500-$2,000/month
Total 2026 annual budget	Burned in 4 months
Reason	Adoption rise + agentic workflows consume tokens

"Burned through entire 2026 AI budget in 4 months." — Uber CTO

→ Lesson: Agent coding cost isn't free. Cost-control is critical skill (section 08).

06 Sub-agent orchestrator-worker pattern

Standard 2026 pattern

┌───────────────────────┐
│  ORCHESTRATOR AGENT    │ (Sonnet 4.6 / Opus 4.7)
│  (Lead, plan, synth)   │
└──────────┬────────────┘
           │
   ┌───────┼───────┐
┌──▼──┐ ┌──▼──┐ ┌──▼──┐
│ Sub │ │ Sub │ │ Sub │   (Haiku 4.5 — cheap, fast)
│ #1  │ │ #2  │ │ #3  │
└─────┘ └─────┘ └─────┘

How it works:

Orchestrator receives big task
Breaks into 3-5 parallel subtasks
Spawns sub-agent (Task tool) per subtask
Sub-agent has fresh context window (no bloat)
Sub-agent returns summary to orchestrator
Orchestrator synthesizes → final output

Anthropic Research case (Apr 2025):

Beat single-agent Claude Opus 4 by 90.2% on internal eval
Cost: 15x tokens vs single chat

Code example (Claude Code SDK)

typescript

import { Anthropic } from '@anthropic-ai/sdk';
const client = new Anthropic();

async function leadAgent(task: string) {
  const plan = await client.messages.create({
    model: 'claude-sonnet-4-6',
    system: 'Break tasks into 3 parallel subtasks.',
    messages: [{ role: 'user', content: task }],
  });
  
  const subtasks = parsePlan(plan);
  
  const workerResults = await Promise.all(
    subtasks.map(subtask =>
      client.messages.create({
        model: 'claude-haiku-4-5',
        system: `Complete this subtask: ${subtask}`,
        messages: [{ role: 'user', content: subtask }],
        tools: [readFile, grep, bash],
      })
    )
  );
  
  return client.messages.create({
    model: 'claude-sonnet-4-6',
    system: 'Combine subtask results into report.',
    messages: [{
      role: 'user',
      content: `Task: ${task}\n\nResults:\n${workerResults.map(r => r.content[0].text).join('\n\n---\n\n')}`
    }],
  });
}

07 5 practical patterns

Pattern 1: Code Review

Orchestrator: review PR → Sub 1 (style/lint) + Sub 2 (security) + Sub 3 (test coverage) → synthesize comment

Pattern 2: Ticket triage

Orchestrator: triage 50 GH issues → Sub 1 (classify) + Sub 2 (priority) + Sub 3 (assignee) → spreadsheet

Pattern 3: Doc gen

Orchestrator: API doc → Sub 1 (extract routes) + Sub 2 (parse types) + Sub 3 (gen examples) → OpenAPI spec

Pattern 4: Migration

Orchestrator: React 17 → 19 → Sub 1 (imports) + Sub 2 (deprecated API) + Sub 3 (tests) → PR

Pattern 5: Research

Orchestrator: "Why slow checkout?" → Sub 1 (DB profile) + Sub 2 (network) + Sub 3 (frontend) → root cause

08 Cost control — 5 levers

5 ways to reduce Claude Code cost

1. Prompt caching (Anthropic native) — 90% input cost off if prefix cache hit 2. Batch processing — 50% off, 24h SLA 3. Haiku for sub-agents — 5x cheaper than Sonnet 4. Context discipline — Grep before Read, sub-agent fresh context 5. Eval gate — test 10 samples before running 1000

Cost cheat sheet (May 2026)

Model	Input ($/MTok)	Output ($/MTok)
Haiku 4.5	$1	$5
Sonnet 4.6	$3	$15
Opus 4.7	$5	$25
Opus 4.7 Fast	$30	$150 (6x)

→ Rule of thumb: 1 typical Claude Code task ~$0.10-2 depending on scope.

09 MCP integration (brief — see Chapter 6)

Claude Code is MCP-native. Setup 1 MCP server:

json

// ~/.claude.json
{
  "mcpServers": {
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": { "GITHUB_TOKEN": "ghp_..." }
    }
  }
}

→ Claude Code uses tool mcp__github__* etc. directly.

→ Details: Chapter 6 — MCP

10 Global freelance with Claude Code

Freelance day rates

Region	Rate/day Claude Code dev
France (benchmark)	€550-900/day
US	$600-1,500/day
UK	£400-700/day
Germany	€500-800/day
Singapore	S$500-1,000/day

→ Emerging market dev knowing Claude Code can bill $200-500/day freelance (vs $30-80/hr local rates).

Skills needed

English communication — required
Claude Code + MCP proficiency — differentiator
Codebase navigation — Grep/Glob/Read fluency
Architecture sense — when orchestrator, when single agent
Cost awareness — don't burn client budget

Where to find gigs

Toptal — top freelance, vetted
Arc.dev — remote-first
Lemon.io — Vietnam-friendly
WIP.co, IndieHackers — solo founder hire
Twitter / X — direct outreach

11 Practice exercises

✍️ 3 levels

Level 1 — 1 week

Setup Claude Code Pro ($20)
Pick 1 OSS repo → contribute 3 PRs with Claude Code
Test 2 MCP servers (GitHub + Postgres)

Level 2 — 1 month

Implement orchestrator-worker for 1 internal task
Measure cost vs time saved
Write tutorial / case study

Level 3 — 3 months

Land 1 freelance gig $200+/day with Claude Code
Build 5 internal MCP servers for team
Eval framework: measure agent task success rate

12 Continue reading

Final word

"Claude Code doesn't replace devs. It replaces devs who don't use Claude Code.Like IDEs didn't replace programmers in 1990. But programmers in 2026 who don't use IDEs = dead.2030 will be devs who don't use agents = dead."

Chapter 2 — Claude Code Deep ​

01 What is Claude Code? ​

Architecture ​

Pricing (May 2026) ​

02 Sonnet 4.5 — "30 hours autonomous" moment ​

Story ​

Why breakthrough? ​

03 Sonnet 4.6 + Opus 4.7 (May 2026) ​

04 Case 1 — Jaana Dogan / Google Gemini API team ​

05 Case 2 — Uber CTO burns AI budget in 4 months ​

Story ​

06 Sub-agent orchestrator-worker pattern ​

Code example (Claude Code SDK) ​

07 5 practical patterns ​

Pattern 1: Code Review ​

Pattern 2: Ticket triage ​

Pattern 3: Doc gen ​

Pattern 4: Migration ​

Pattern 5: Research ​

08 Cost control — 5 levers ​

Cost cheat sheet (May 2026) ​

09 MCP integration (brief — see Chapter 6) ​

10 Global freelance with Claude Code ​

Freelance day rates ​

Skills needed ​

Where to find gigs ​

11 Practice exercises ​

12 Continue reading ​

Chapter 2 — Claude Code Deep

01 What is Claude Code?

Architecture

Pricing (May 2026)

02 Sonnet 4.5 — "30 hours autonomous" moment

Story

Why breakthrough?

03 Sonnet 4.6 + Opus 4.7 (May 2026)

04 Case 1 — Jaana Dogan / Google Gemini API team

05 Case 2 — Uber CTO burns AI budget in 4 months

Story

06 Sub-agent orchestrator-worker pattern

Code example (Claude Code SDK)

07 5 practical patterns

Pattern 1: Code Review

Pattern 2: Ticket triage

Pattern 3: Doc gen

Pattern 4: Migration

Pattern 5: Research

08 Cost control — 5 levers

Cost cheat sheet (May 2026)

09 MCP integration (brief — see Chapter 6)

10 Global freelance with Claude Code

Freelance day rates

Skills needed

Where to find gigs

11 Practice exercises

12 Continue reading