A book by Michael Elkær Meisner
The Agent Doctrine
Running AI agents the way we run everything else that matters.
An agent is a worker without a contract, and everything we know about running workers, suppliers and systems applies to it. The book follows one organisation from the day it discovers the problem to the day its board can see the answer.
Get the book
Free
The PDF
No sign-up. Share it with your board, your team or your suppliers.
Available at releasePaperback
On Amazon
For the shelf and the meeting table.
Available at releaseThe book will be licensed under CC BY 4.0: copy, share and adapt it, including the ten rules and the figures, with credit to the author and the title.
The ten rules
- Every agent has an owner and a purpose. One named person answers for it, and one sentence says what it is for. If nobody can be named, the agent does not run.
- Every agent has an end date. On that date it is renewed or retired. Silence no longer means that it continues.
- An agent gets no more access than its job needs. It has a known identity, and its reach is something someone decided.
- The human moves up the stack. People set the limits, receive the exceptions and take the decisions an agent may not. The limits are held outside the agent.
- Control follows size. How much governance an agent gets depends on what it can reach and what happens when it is wrong.
- Nothing runs without a record. If the question "how many agents do we have, and what can they do" takes more than a day to answer, there is no governance yet.
- Proof before trust. A demonstration is a handful of observations. Evidence is many cases with known answers, run again after every change.
- Change is a decision, not an event. New data, new permissions, a new model or a new purpose are decided by someone, or found and tested when they arrive from outside.
- An agent can fail, be stopped and be accounted for. Incidents are recorded, a named person can stop it, and a log shows what it did.
- A vendor's agent is a vendor relationship. Data, scope, price, model changes, liability and exit are settled in writing before the agent matters.
Who it is for
Board members get ten questions to ask. Those who have to make the answers true get a method, a ladder to measure it by, and a practical guide in the operational appendix.
How this book was written
With Claude, made by Anthropic, under the author's direction and responsibility. The book applies its own Doctrine to the way it was made, and a note on authorship describes how.
Templates that improve with use
The checklists and templates in the operational appendix will be updated as the Doctrine is used. A sign-up for new versions opens with the book. Until then, follow the release on LinkedIn.