the weekly agent lessons, and what changed
For beginners: the four-station loop that turns a chat model into an agent, written out by hand in JavaScript and in Python. Why the model never runs anything, why tool descriptions are prompts, and the four rules that save you a weekend — then an exercise you run in the tab.
A real language model running on your own GPU, inside the browser, with the network off. What a model actually receives and sends back, a complete single-file WebLLM example, the web worker version, both ways to give it tools, and what “free” costs you in download and VRAM instead of money — and why a weak model is the right one to learn on.
The course has a shape: tool calling, loops and goals, agents hiring other agents over Blocks.ai, then a model in the tab with WebLLM, vision and evals. The loop that makes an agent an agent comes first, with what you build and what you need for each — and an argument for why that order and not another.
buttercup.sh is now a course as well as a harness. One lesson by email every week — a single idea, the code that makes it work, and an exercise you run in the tab. Free, no account, and every exercise runs on a local model if you would rather not spend a cent.
Eighteen tools, a virtual filesystem and an undo stack in one static page — on desktop or phone, with no backend and no build step. Free forever, and it talks to local models so nothing has to leave your machine.