Guide Book

A reference you can read front to back or dip into. Every chapter explains the syntax, and most topics have a 🔍 Behind the scenes section you can open to see what Python is actually doing, the details classes usually skip. Look for 🧭 Scenario cards for real-world uses, and press Try it on any example, then open the Inspect tab to look inside.

1. Chapter 1What an agent actually isA model, a tool list and a loop — the whole idea, and the three stop reasons that drive it.
2. Chapter 2ContextThe window is the agent's whole world — what fills it, what it costs, and the four ways to keep it from overflowing.
3. Chapter 3Designing toolsA tool is an interface for a reader who has never seen your codebase — naming, schemas, error messages, and how many tools is too many.
4. Chapter 4SkillsInstructions that load on demand — the frontmatter contract, how one gets picked, and what makes a skill worth writing.
5. Chapter 5MCPOne protocol between a host and a capability provider — servers, transports, the three primitives, and where the trust boundary sits.
6. Chapter 6The harnessEverything around the model — CLAUDE.md, settings and permissions, hooks, and why a harness is mostly about what the agent is *not* allowed to do.
7. Chapter 7Prompting an agentWhy prompting an agent differs from prompting a chatbot — goals over procedures, verification loops, and the patterns that stopped working on newer models.
8. Chapter 8Subagents & delegationWhen a second agent earns its cost, what it can and cannot pass back, and the orchestration patterns that actually hold up.
9. Chapter 9Prompt injection & trustEverything a tool returns is data, not instructions — why that line is the whole security model, and what to build around it.
10. Chapter 10Evaluating agentsHow to know whether a change helped — building an eval, grading it, and hill-climbing without fooling yourself.