Agents Are Easy Governance Is Hard

Agents Are Easy. Governance Is Hard.

**A working agent takes an afternoon. Getting it allowed to run takes the rest of the quarter.**

This repo is for the person who has just been asked to "do something with AI agents" and wants to know what they're actually signing up for. It covers the three ways to work with Claude — **Cowork**, **Claude Code**, and the **Agent SDK** — when each one is right, and the part every tutorial leaves out.

The centerpiece is two examples that do **exactly the same thing** — byte-for-byte identical output. One is 40 lines of executable code. The other is 325. That difference is the entire subject of this repo.

---

The one-paragraph version

Getting a language model to read a document, decide something, and take an action is genuinely easy now. You can do it in an afternoon, and it will work, and you will be pleased. Then you try to put it somewhere real — where it touches customer records, or spends money, or sends things to people — and you discover the actual job. Who approved this? What can it reach? What happens when it's wrong? How would anyone know? What did it do last Tuesday, and can you prove it?

None of those are model problems. All of them are engineering problems, and they're the ones nobody blogs about.

---

Start here

**Never used any of this?** → [Your First Hour](docs/02-your-first-hour.md) — from nothing to a working agent, with the honest version of what you just built.

**Deciding which tool?** → [Cowork vs Claude Code vs Agent SDK](docs/01-cowork-vs-code-vs-agents.md) — a real decision guide, not a feature matrix.

**Already have a demo and need to ship it?** → [The Governance Gap](docs/03-the-governance-gap.md) and the [Governance Checklist](docs/04-governance-checklist.md).

**Want to see the difference, not read about it?** → [`examples/`](examples/) — the same agent, twice.

---

Three ways in