Harness Engineering

The thesis. Why a model alone is not an agent, the precise difference between prompt, context, and harness engineering, and the five layers you will build across the workshop into your own coding harness.

Day 1. The beating heart of every harness: a loop that calls the model, runs the tools it asks for, and repeats until the work is done. We build a bare harness from first principles.

Day 2. A harness that can only talk is a chatbot. We give it hands: file and shell tools with strict schemas, permission gates before dangerous actions, and a sandbox to contain the blast radius.

Day 3. The context window is the scarcest resource in the system. We build the machinery that decides what the model sees every turn, compacts long sessions, and remembers across them.

Day 4. Real agents crash, get interrupted, and hit jobs too big for one context. We make the harness recoverable and self-healing, and teach it to dispatch sub-agents under supervision.

Day 5. We dissect three real harnesses to see how they implement, at scale, everything you just built by hand, then you assemble your own and demo it.

How Pi actually works, built up from an empty file the way you would build any coding agent, one layer at a time — but every layer is Pi's real code (earendil-works/pi, MIT). We start with the loop and add exactly one thing per chapter, each because the last chapter left something missing: the toolbox, tool safety, the system prompt, the execution environment, context and cache, subagents, surfaces, and the extension system. Every claim is read straight from the source.