An accurate picture of the tool before any habit is asked of you. Six topics trace three generations of coding assistant to the agent that edits files and runs commands, give the loop it runs in one paragraph, and map what it knows, what it can touch and where it runs, before you meet the codebase the rest of the book works on.
6 topics
Most of the trouble engineers have with coding agents comes from a wrong picture of what the agent is. They treat it as a smarter autocomplete, and stop reviewing. They treat it as a colleague with a memory, and are surprised every morning. They treat its "done" as a test result, and merge on it. Each of those pictures is close enough to work on a good day and wrong in a way that costs a week on a bad one.
This chapter replaces them with an accurate one. It starts from what changed between editor completion, chat and agents, which is not the model but who decides the next step. It gives the agent loop in one paragraph and names the parts you supervise. Then it covers the three facts that explain most of an agent's behaviour: it knows only what is in its context this session, it can do anything its tools and your account can reach, and the place it runs decides how far that is.
The chapter ends at Cragmoor, a climbing-gym company with a three-year-old booking service, the team you have just joined and the first task you will be given: cut the full-refund window from 24 hours to 12. No habit is asked of you yet. The habits start in Chapter 2, and every one of them follows from what is on these six pages.
What you are working with, layer by layer
Where it runs
decides how far its tools reach, and how much of the run you see
What it can touch
a short list of tools, with permissions deciding which ones ask first
What it sees
its context this session, and nothing about your project beyond it
The loop
a model that picks its own next step until it judges itself done