I've had post titled "Take $200 and open your eyes" sitting in my drafts for weeks now. Might be time to write it now, here's what it boils down to: 🥚This is new technology. Don't close your eyes. Open them, get close, see for yourself what's real and isn't. 🥚If you last tried AI or "agents" in Cursor half a year ago, your knowledge is outdated. 🥚If you use the free version of anything, your knowledge is useless to assess what's possible on the frontier and what's coming. If you use local modals on your machine, your knowledge is useless. If you haven't used Opus 4.5 or Gemini 3 for programming, your knowledge is useless. 🥚Yes, this will cost money. Yes, it's not cheap. Yes, it's worth more than anything anyone writes about AI anywhere. If you've been programming for >1 you can't conceptualize, without practice, what these agents can and can't do. You try, but practice will smash your theory. 🥚Take money. Take $200, go use Amp, or Claude Code, or even Codex. Make it your goal to spend those $200. Don't save it up: *this is the exercise*. 🥚This is the MOST IMPORTANT BIT: do not form any opinion until you've spent the $200 in tokens. Your goal is to LEARN, to be curious, to see, to understand, to feel. Do NOT form opinions. You can only have opinions once you've spent the $200. And even then: you don't need to have opinions. 🥚Do NOT show up in the first session and try to "gotcha!" the agent. You are here to learn. VERY proficient programmers (Armin is one of them) are getting A LOT out of these tools. When the agent doesn't do what you think is good, don't go "ha!! I told you! They can't write code!" but try to think "wait, what am I missing, why is it working for Armin? Why is Thorsten saying that he rarely types code by hand anymore, assuming he isn't a complete idiot?" 🥚Pick a challenging engineering project and think through how you would build it. Put that in the prompt. Start throwing Amp or CC at it. 🥚OBSERVE! What is it doing? Which tools is it using? Can I give it more feedback somehow? How can I tell the agent to check its work? How can I put within reach of its tools the things it needs to know to decide it's on the right track? Build feedback loops. 🥚Try to produce good code! Do NOT "vibe code": do not check out and not look at the code. You haven't spent enough in tokens yet to do that. Try to do good engineering! Guide the agent to do it for you! Let the agent do the typing and you try to do the thinking and then, later, you both do it. 🥚Write a web app, a good one. Write a CLI, write a text editor, write a terminal emulator, write a shell, a small parser, a small interpreter. Connect it to a SQL database by telling it to use psql, have it write reports, build a dashboard. Throw it at a complex codebase and try to fix a bug or add a feature. GO WIDE! YOU ARE HERE TO LEARN! THESE TOKENS ARE LEARNING BUDGET! 🥚Austin Henley's list of challenging engineering project. Pick one. Build it with the agent. 🥚 While you're doing this, see what the agent is repeatedly doing wrong or missing. Add it to the AGENTS.md. See what tools it's missing in the codebase: have it add those tools. You want the agent and codebase to become a single entity. 🥚 Keep thinking: what else CAN it do? what else could I make it build, now that I don't have to type anymore? Tests, fuzzers, little CLIs and helpers, tooling. Build it! 🥚 Forget about multi-session stuff for now. Try to use one agent and look at it and its output. 🐣 Spent the $200? Now you can start to form opinions. Try not to though.