💡 This course is agent-agnostic. Concepts work with Claude Code, Codex, Hermes (the open-source agent the author uses in Telegram), and the open agents the author ran on the same tasks in an isolated sandbox: goose, OpenCode, Pi, nanobot, CheetahClaws, dsh (DeepSeek), and Unreal Agent. If you work from China or Russia, behind a censorship firewall (the Behind-Firewall stack), or want a sovereign stack on your own hardware (Sovereign) — module 03 ("Stack selection") shows 4 options.
<aside> 🧭
Interactive block — on the site: open →
</aside>
💭 in your head · nothing to type
Picture the last time you asked something from AI — and got back something close, but not quite right.
Don't analyze it as a "bad prompt." Instead, picture yourself in a restaurant where the menu is half in a language you only partially know. You guess at words, point at a line, the waiter nods — and brings something adjacent to what you meant. No one's fault. Just an untranslated word sitting between you.
Run that moment in your head: where exactly did the untranslated word sit between your thought and the AI's answer? Catch it — hold it there.
💭 in your head · nothing to type
Now shift the frame. Not the restaurant — a map printed in someone else's handwriting. You know where you want to go, but half the labels are smudged. The AI navigates by the map you handed it.
Notice in your head: which landmarks on your map were sharp — and which were missing? Where did the AI invent a route instead of reading one?
Most people catch the same thing here: the problem isn't the navigator — it's the map we hand over.
This framing comes from Andrej Karpathy: he described Software 2.0 in 2017 and Software 3.0 in a 2025 talk.
Software 1.0 — we wrote code by hand, line by line. The computer did exactly what we said.
Software 2.0 — machines started learning from data. Neural networks. We give examples, not rules.
Software 3.0 — agents execute code from our intent. That's the tool: what it can do. What you do is a separate question. Not "AI writes instead of me," but co-thinking and co-working: you and the agent think and build together. And there's more than one way to co-think. Sometimes you lead step by step, sometimes you share the wheel, sometimes you hand over only the goal.
| Before | Now |
|---|---|
| I write code → machine runs it | I set the intent → co-work with the agent → the decision stays mine |
| I solve EVERY problem myself | I choose the way of co-thinking that fits the task |
| The tool works for me | The tool thinks and builds with me |