Projects
Reading is not enough — a harness is learned by building it. Each project is one layer you build with your hands, pointed at the exact chapters that carry it. Do them in order across the five days, and by the end you have your own working coding agent.
The bare loop
Build the smallest thing that is genuinely an agent: a model client, a message array, and a hand-rolled loop that calls the model, runs the tool it asks for, and repeats until the work is done.
Give it hands
Wire real read / write / edit / bash tools into the loop with strict schemas, then add a permission gate and a sandbox so your harness edits actual code on your machine, safely.
The context engine
Add compaction plus a persistent memory layer so your harness survives a 200-turn session without losing the plot and starts each run already knowing the project.
Make it durable
Checkpoint every model turn and tool call so a crashed session replays instead of re-running, then add self-healing retries. Kill the process mid-run and watch it resume.
Dispatch sub-agents
When one context can't hold the job, spawn a focused sub-agent, let it work, and return only its conclusion — behind a supervisor and a human-in-the-loop approval gate.
Your own harness (capstone)
Assemble the loop, tools, context engine, durability and orchestration into your own pi-style coding agent, evaluate it, and demo it live. Yours forever.