The Agent Playbook is out: agents that finish work, and the discipline that makes them safe
Spend any time around AI agents right now and you'll notice a pattern: everyone has a demo, almost nobody has a deployment. The demos are dazzling — an agent researches, writes, posts, buys. Then someone asks the questions that matter. What happens when it's wrong? How do you stop it? How do you know it's still working in three weeks? And the demo goes quiet.
I kept running into that gap in my own work. I build and run AI systems every day — Archon, an open-source control room for agent fleets, and a live-validated forex trading system that runs unattended with hard stops, weekly honest reviews, and a scoreboard that doesn't lie. The trading system taught me something the demos never do: a system you can't measure, bound, and stop isn't a system. It's a liability with good marketing.
So I wrote The Agent Playbook — 41 pages on building, controlling, and shipping AI agents that finish their work. It's out today at £29, and this post is the honest story of what's in it and why.
The thesis: finishing is the skill
An agent is a loop: perceive, reason, act, observe, repeat. Anyone can start the loop. The engineering is in making it stop — at the right moment, for the right reasons, with verifiable output. That means stopping conditions designed up front. Tools granted by least privilege. Approval gates on anything consequential. Budgets that turn a runaway loop into a bounded failure. A kill switch tested like a fire alarm — on a schedule, not after the fire.
None of that is glamorous. All of it is the difference between a demo and a service. The book's whole argument is that this "boring discipline" is where the value lives — and that most of it transfers from fields that learned it decades ago: aviation, finance, operations.
What's inside
- The automation filter — five questions that save you months: frequent? cheap to be wrong? verifiable? well-specified? fat middle? Most "agent projects" fail this filter, and knowing that early is the cheapest insurance in the field.
- Designing agents that finish — goals, boundaries, stopping conditions, and the human's real role: defining and verifying, not rubber-stamping.
- The control stack — least privilege, approval gates, rate limits, spend caps, and the kill switch, layered so each catches what the last missed. Plus the attack surface: prompt injection, exfiltration through tools, and why structural boundaries beat instructions.
- Measuring value with real numbers — completion, rework, cost per finished task, cycle time, escalation rate — and a worked case with actual figures: £485 a week of human triage down to £148, £17,500 a year saved, with the trajectory to prove it.
- Shipping — shadow mode, limited live, evaluation suites as regression tests, the model-update problem, and a nine-line checklist you can steal.
Who it's for
Chapters 1–9 are for everyone: if you work with AI in any form, they'll change how you think about delegation and trust. Chapters 10–14 get hands-on — prompts, tools, eval suites, deployment — where some engineering fluency pays off. If you're a developer, a technical lead, or the person in your team who keeps saying "wait, but how do we know it did it right?", this is your book.
Try before you buy
I'd rather show than tell, so Chapter 8 — Choosing What to Automate — is free. It's the chapter people will get the most immediate use out of: five questions, five scored tasks, and the honest list of what should never be an agent. Download the sample chapter (PDF) — no email needed.
The buy
The Agent Playbook — £29, instant PDF, 41 pages. No subscription, no upsell inside, no fluff. If it saves you one month of building the wrong thing, it's paid for itself many times over. And if it doesn't land for you, email me — I'd genuinely rather know.
— Akeem · keemsai.com