Build a real AI agent from scratch, one tiny piece a week.
Open a blank file. Each episode adds one small, readable thing until it talks, remembers, texts you, and acts on its own. No framework, no magic. Just the code.
Seven episodes. Each one ships code, a Short, and a write-up.
It starts as 12 lines that can talk. By the finale it remembers you for weeks, messages you first, and can swap its own brain mid-conversation. Read it top to bottom and you have built the whole thing.
It talks — the agent loop
No framework, no server. The whole core is one function: prompt in, answer out.
It remembers — SQLite
The model forgets everything between messages. ~20 lines of SQLite give it a memory that survives restarts.
It texts you — WhatsApp
About 20 lines and the claw moves out of the terminal into WhatsApp. Text it, it texts back.
It gets hands — tools
Three tools and it can act: open a browser, search your files, send one back. "Send me my quarterly report" and the PDF lands in chat.
It has a heartbeat — proactive nudges
A timer that messages you first, nudging you to drink water every few hours and tracking how much you have had.
It keeps a diary — long-term memory
Durable facts. Tell it once (an allergy, a goal); weeks later, in a fresh chat, it still knows.
“Every agent tutorial I found started with a framework and a thousand lines of boilerplate. So I deleted all of it and rebuilt the thing from a blank file, small enough that you can actually read every line and know what it does.”
— Chandra, building in public on SennaMind
