(Ch. 0.)
“No, Sir, do you read books through?” Samuel Johnson, quoted in James Boswell, The Life of Samuel Johnson, LL.D. (1791)
Something is on fire at your place this week. Maybe an agent reported the ticket closed and the ticket was not closed, and you found out from a customer. Maybe your token bill tripled on a Tuesday and nobody changed anything, which is a sentence that has never once been true. Maybe you asked for a migration and got twelve numbered steps with risk callouts, beautifully formatted, and zero commits. Maybe it fetched a URL some stranger put in a support ticket, from inside your network. You did not sign up to manage this thing. It showed up in your org chart anyway, and it reports to you.
Bottom line: This is an onboarding packet, not a course. You got promoted to supervisor of something smarter than you and faster than you and occasionally full of it, and nobody handed you the manual, so here it is: 62 short chapters, each one a pattern that shows up when an agent has to do real work and not lie about it. Every chapter stands on its own. Bottom line, when it bites, the pattern, one real story from production, the quiet failure, do and don’t, receipts. There is no prerequisite reading. There is no build-up. Open the one that is on fire, fix your thing, close the book, go home. If you read it front to back anyway, that is fine, but you are doing it for fun and we both know it.
Part 0, front matter (Ch. 0-7, plus the Interlude). Where this came from, why the AGI argument is over, why the bubble is financial and not technical, and why your fine-tuned little model is going to embarrass a $200/month frontier subscription at your specific job. If nothing is on fire, start at Ch. 6 and find out whether the thing you built is actually an agent.
Part I, talking to the damn model (Ch. 8-13). Prompts, context, personas, formats, skills. If your outputs are inconsistent and you have been rewriting the same mega-prompt for a week, open Ch. 12. It is the context, stupid.
Part II, the loop (Ch. 14-20). The outer loop, planning, routing, tools, self-critique, stop conditions. If your agent narrated a plan instead of executing it, that is Ch. 15, and it is the most-linked chapter in this book for a reason.
Part III, knowledge (Ch. 21-24). Memory, retrieval, MCP, and the citation it made up for a fact that was already true. If it cited a source and the source does not exist, Ch. 23, and go open the file.
Part IV, many (Ch. 25-30). Parallel work, roles, model routing, agent-to-agent contracts, and the merge. Fan-out is the easy half. If your parallel run produced six branches nobody can combine, Ch. 27.
Part V, reliability, which is why this book exists (Ch. 31-40). Ten chapters on the gap between “it said success” and “it worked.” If you only ever read one part of this book, read this one, and start at Ch. 33: it said success and did the wrong thing.
Part VI, safety (Ch. 41-43). Hard stops in code, not moods in a system prompt. If your agent can reach your LAN, Ch. 42, today, before lunch.
Part VII, ops (Ch. 44-51). Metrics, the bill, clustering your own boxes, logs you can read at 3am, and the cron that died four months ago while every dashboard stayed green. Bill tripled? Ch. 45. Everything looks up and nothing is happening? Ch. 48.
Part VIII, getting better (Ch. 52-55). Evals, skills, open source as a door-opener, and why the brilliant jerk is now optional. Start at Ch. 52: yelling at it does not update the weights.
Part IX, go deep first (Ch. 56-61). The pstack plays. Prove it, subtract before you add, build the lever. Start at Ch. 56, and this is the one part that rewards reading in order.
Then Ch. 62, which belongs to no part. The one where I stripped a local model’s ability to refuse anything, told it four hundred times that the other models were beating it, felt fine about that for three days, then described my own Tuesday out loud at a dinner table and watched everyone stop chewing. It sits at the back because that is where it goes.
Same shape, every time. Bottom line first, so you can bail after one paragraph and still have gotten your money’s worth. When it bites, so you can tell whether this is your fire. The pattern, with whatever the actual research says. One worked example from production, not a toy, with dates and commit hashes. The quiet failure, which is always the expensive one. Do and don’t. Receipts.
Three rules I do not break.
No invented citations. Every number in here has a source behind it. Where I could not verify something, I either cut it or told you in the sentence that it is my memory and not a measurement, so you know exactly how much weight to put on it. I would rather say “I do not have the receipt for this one” out loud than fake a footnote, and I would rather hand you an honest under-construction GIF than a confident lie with a URL attached. Chapter 23 is about agents doing exactly this to you. It would be a bad look to do it back.
I am the idiot in the stories. Not a composite
engineer, not “a team I worked with.” Me. I wrote the
continue statement that threw away the evidence, then spent
a day debugging my own good judgment. I put a number in a P0 that was
off by 3x because I multiplied by the wrong price. Twenty-five years of
shipping software, pwned by one line of arithmetic.
The robot on the cover is named Clark. He is the agent I run everything through, he wears my logo, and when I asked him who he thought he was, he told me he was my alter ego and sent a picture. The picture is in Chapter 21. I did not approve the Speedo.
Her name is Georgia, and she likes the machine more than she likes me. See the Interlude. It is two pages, it is true, and it is the only chapter with no do/don’t section because there is nothing I can do about it.
Not a framework tour. LangChain, LangGraph, Crew and ADK get one honest page in the back about when you would actually reach for each, not a shrine. Not a course, because there is no curriculum and no certificate and no capstone project. Not a prediction. I am not telling you what 2029 looks like, I am telling you what broke in production last August with the timestamps still attached. And not neutral. I have opinions, they cost me money to acquire, and hedging them into mush would waste both our time.
Your new report has already read your email. Least it could do is read this too.
Verified