Your First AI Automation: Let Your AI Work While You Sleep — Safely, and Without Writing Any Code
Building an AI that acts is one thing. Letting it act when you're not there is another - and it needs a different kind of owner.
In a couple of evenings you'll give your AI agent a schedule. It wakes at seven, checks a few things you care about, decides whether any of it is worth your attention, and sends you a short brief. You will not type a single line of code, or a single terminal command, at any point.
Here's the part most books about automation won't tell you: most mornings, it sends nothing at all. That's the version that works. An automation that speaks every day isn't a briefing - it's a subscription you'll cancel.
This is the third book in the series, and it stands alone. If you've read Your First AI Agent, the ten moves carry over and Chapter 3 takes four minutes. If you're new, Chapter 3 builds you an agent from scratch. But be warned about one thing either way: you cannot safely schedule something you have never watched run, and this book will make you run it by hand five times before it lets you near a timer.
Why does leaving the room change everything? Because every safeguard you'd normally rely on quietly assumed you were there. A live activity feed is a tree falling in a forest at 4am. A confirmation dialog that stops and asks has only two outcomes when nobody answers: the job hangs forever, or - far more likely - you delete the dialog on a morning you're in a hurry. A limit on one task says nothing at all about the same task running three hundred times.
So the job becomes four new things, and they're what this book is really selling. BRIEF: write standing orders, because nobody will be there to ask - and the test for every line is whether past-you would defend it to present-you at 7am on a bad morning. RECALL: give it a record of what it already did, or it will do the same job forever. DIGEST: supervision goes asynchronous, in a form short enough that you actually read it. BUDGET: a limit bounds an event, but a schedule creates a rate, and a one-cent job is $3.65 a year daily or $5,256 a year every minute - the same job, one character apart.
You'll also break it on purpose. A whole chapter has you sabotage its memory four ways: no memory at all, memory it never consults, memory that only records its successes, and memory that never forgets and quietly gets more expensive every morning. The lesson is calmer than the headlines: not one of those is an AI deciding something alarming. All four are bookkeeping failures - and almost every automation horror story you've read is one of them wearing a costume.
And one chapter is about what you should never hand over at all. Four tests, the draft-don't-send pattern that rescues most tempting cases, and a requirement that you take something back off your agent before the last page. The book isn't finished until you have.
Inside: what a schedule actually is, in plain English - the timezone trap that catches everyone - a complete worked set of standing orders, annotated - the one outbound power, and why you make it safe by making it narrow rather than by trusting the model - alerting on silence, because a quiet agent and a dead one look identical from a phone - the Friday review, and the four design decisions that make skipping it survivable - deploying something that might run mid-edit - plus a prompt pack of all fourteen moves and a symptom-to-sentence table.
Requires a paid Claude plan and a few dollars of API credit, both stated plainly in Chapter 1. No experience of any kind assumed.
By the last chapter something you built runs while you sleep - and you can say exactly what it was allowed to do, what it did, what stopped it, and what it costs to keep.
Customer Reviews
ThriftBooks ® and the ThriftBooks ® logo are registered trademarks of Thrift Books Global, LLC
