A 90-SECOND PLAYABLE LESSON

One AI brain vs an orchestra

Feed twelve real jobs to one giant model and watch its memory — and your money — disappear. Then run the same jobs the way agentic engineers actually do.

Sound on (there's a mute button). Two rounds, one receipt.

Prefer reading? The full lesson is below ↓

The lesson, as text

Why one giant AI brain loses to an orchestra

Give one frontier model — Claude Opus 5, $5.00 per million input tokens — a real job: build and ship an app in 12 tasks. Planning, reading the whole codebase, building features, debugging a failed deploy, translating the UI, analyzing usage data.

Everything it reads piles into one context window, and every new task re-reads the whole pile. By task 10 the window is past a million tokens: out of memory, $31 burned, 5.7M tokens processed, nothing shipped.

Same model, same 12 jobs, one change: the frontier model only directs. A crew of 16 Claude Haiku 4.5 workers ($1.00 per million tokens) does the reading in separate, disposable contexts — big jobs get sharded across several workers in parallel. The orchestrator never holds more than briefing notes and one-page summaries.

All 12 jobs ship for $3.07 — 10× cheaper — and the big context tower never fills. That is the whole lesson: in agentic engineering, context is the real cost. The expensive brain should think; cheap hands should read.

The twelve jobs

  • 1. Plan the app 12K tok
  • 2. Read the whole codebase 350K tok
  • 3. Design the database 10K tok
  • 4. Build signup & login 40K tok
  • 5. Build the payments page 55K tok
  • 6. Wire up the AI tutor 35K tok
  • 7. Write the tests 30K tok
  • 8. Debug the failed deploy 220K tok
  • 9. Translate it into 7 languages 60K tok
  • 10. Analyze the usage data 180K tok
  • 11. Review every change 90K tok
  • 12. Write the launch notes 8K tok

Published API prices, no caching or batch discounts, as of August 2026.

This is lesson #1 of the playbook

I'm turning what building a seven-app network with AI agents taught me into the multilingual AI-GTM playbook. The waitlist reads it first — free.