What 3 Months of Running an AI Agent on a Raspberry Pi Taught Me About Task Design

I've been running an AI agent on a Raspberry Pi 5 since May. Not a chatbot, not a demo, an actual agent that manages email, writes code, deploys software, and submits articles. The model running locally is a 3B parameter quantized LLM. This post is about what I learned the hard way about designing tasks for small models. The setup details are covered in another post. Here I want to talk about the part nobody writes about: what happens after you get the model running and try to make it do real…