How to turn a coding agent — Claude Code, Codex, Pi, or any of their cousins — into a delegated operator that runs your real work against external systems, through a documented, authenticated tool belt of small CLI scripts.
A coding agent is very good at reading instructions, calling tools, and reasoning about what to do next. If you give it a set of small, well-documented command-line tools that talk to a real system as you, your workstation becomes an automated operator that does your work — not a chatbot that talks about it.
This guide builds that idea from first principles to a concrete reference implementation. We do not stop at theory: we design a working operator for an eBay seller business, define its tool belt, and write the soul.md that gives it judgment. We deliberately stop short of the tools' source code — the valuable, durable artifacts are the architecture, the contracts, and the operating instructions, not a hundred lines of Python that will be stale by next quarter.
This guide is written in two layers:
Chapters are tagged concept or build so you can read the layer you need.
soul.md, shared skills — and check current API and Terms-of-Service details with the provider before you ship.