The harness. What surrounds the model.

An AI model, on its own, answers. For it to do work it needs everything around it: how it is asked, what it remembers, which tools it has and who reviews what it does. That whole set is called a harness.

The model is the engine. The harness is everything else.

Switching models is switching engines. What keeps the work just as good with one as with another are these four layers.

Instructions

How the work is done and what is never done. The same across every tool.

Skills

Procedures already tuned for what repeats: diagnosing a failure, reviewing a change, planning.

Specialists

Agents with a bounded role: one builds, another reviews, another looks for risks.

Memory

What gets decided and learned is not lost when the conversation closes.

Two pieces. Both open.

The harness started as loose files on one computer and was tuned through daily use. Today it is open source that anyone can install.

• JorgeX StackInstalls and maintains the harness in Claude Code, Codex, OpenCode and Pi with a single command. MIT licence.pnpm dlx jorgex-stack@latestView the code
• JorgeX PiCustom extensions for Pi. The first one groups the agent's activity into collapsible blocks, so the conversation stays readable.View the code