The eight stages of AI coding — and the cost talk
Steve Yegge’s ladder runs from typing into ChatGPT to orchestrating a hierarchy of agents. Knowing where you stand — and where the sweet spot is — is half the game. Then: a straight word about money.
01The ladder, one rung at a time
Steve Yegge — author of the (huge) tool Gastown — wrote a hugely popular post describing the stages you climb as you use AI more and more aggressively. Read it as a self-assessment: which rung do you mostly live on?
Stage 1 · AI autocomplete & ChatGPT
Ask ChatGPT, use AI autocomplete. You copy answers over by hand.
Stage 2 · Agent in the sidebar
A coding agent in your IDE sidebar that asks permission before every change. You approve each step.
Stage 3 · YOLO in the sidebar
Same, but you accept everything. “Go for it, I trust you.” A level of trust given to the sidebar agent.
Stage 4 · Agent takes the main window
The agent fills the screen. Your attention is on the agent, not the code — diffs fly by; you glance, you don’t approve.
Stage 5 · Move to the CLI
Claude Code or similar in the terminal, on YOLO. Diffs scroll past; you sit back, or walk away and come back done.
Stage 6 · A small team of agents
Not one agent — several (say 3–5) running concurrently in your CLI, like a little team.
Stage 7 · Ten-plus agents, you conducting
10+ agents running at once, you coordinating them all. Hugely leveraged, and a lot to manage.
Stage 8 · Agents orchestrating agents
A hierarchy: manager agents running worker agents — testing teams, design teams — orchestrating each other.
In the original post, Yegge had Google’s Nano Banana generate an image for each stage; the eighth — a whole hierarchy of coordinating agents — really captures where the frontier is heading.
Week 1 covers stages 2–4 (IDE agents). Week 2 drills stage 5 (the CLI). Week 3 adds 6, 7 and 8 — orchestrated swarms. But here’s the honest take: for building quality software in the enterprise today, the sweet spot is roughly stages 5–6. Stages 7–8 you’ll see and understand; whether you’d actually run them in production is an open question. Five and six is where the reliable, scalable value is.
02The cost talk
Time for a straight conversation about money — the hottest topic in the field.
You can take this whole course for $0
Not a cent of API cost is required. The resources include tips for setting up free models; where a paid tool appears, take it as a demo for future reference and use a free tool instead.
It’s also easy to spend a lot
Spin up 20 agents to build something enormous and they’ll happily take your money. Sometimes that’s worth it — but it’s your call, made with eyes open.
To get the most out of the course: use the Cursor free trial while you’re eligible, then the $20/month Claude Code plan for Weeks 2–3. Go beyond that only if you want to get more intense. If you’d rather spend nothing, follow the free-model notes in the resources and skip the paid demos.
The course takes responsibility for getting you to “answer the Karpathy tweet” by the end. It cannot take responsibility for your AI spend — pricing varies by region, offers, and trials that change constantly. You’re in the driving seat on cost. You decide what you’re comfortable with and you monitor your usage; the course gives you maximum transparency on where and how.
A reframe on the friction: a $1,000 laptop doesn’t shock us, but lots of $20–30/month subscriptions feels like being nickel-and-dimed. Keep in mind each AI response is hundreds of trillions of floating-point operations — someone has to pay the electricity bill. That there’s any tool this capable is astonishing; some cost is unsurprising.
✓ Key takeaways
- The eight stages run from ChatGPT autocomplete to agents orchestrating agents.
- Stages 1–4 are IDE; 5–8 are CLI. The enterprise sweet spot is 5–6.
- More advanced isn’t automatically better — match the stage to the task.
- You can complete the course for free; a small paid plan is recommended, not required.
- You own your AI costs. Decide deliberately and monitor them.