Building the Full Mini-Harness

Time to assemble. Every MVP from Chapters 2 through 14 was a real part of one system, and this chapter snaps them together into mininharness, a small but genuine terminal coding agent. It runs the loop, builds layered prompts, manages context and budget, runs tools behind an approval gate, loads project memory and skills, and can delegate to a subagent. It is a few hundred lines, small enough to read in a sitting, and shaped like the production systems this guide dissects. The only thing it does not ship with is a real model; you drop in an API call (Anthropic, OpenAI, or a local model) in one place, and everything else is the harness, which, as Chapter 1 promised, is 95% of the work.