Taming the Fire — Keeper in the Loop

Prologue: The New Fire

The machine surfaces; the human decides.

Fire was the first powerful tool, and the whole story of being human is the story of learning to stay involved with it. We did not put it out, and we did not walk away. We built the hearth, set the guard, and tended the flame with a careful hand — never reaching into it — because fire warms a home or burns it down depending entirely on whether someone is paying attention. Agentic AI is the new fire. It can do extraordinary good at extraordinary speed, and it can cause extraordinary harm just as fast, and the difference is the same as it has always been: is a human still in the loop?

This is the part of my agentic journey I most want to be useful to other people, so I am writing it plainly. I am not going to pretend I invented the idea of human oversight — I did not, and you should be suspicious of anyone who claims to. The field already has good words for this: human-in-the-loop, guardrails, approval gates, stop authority. Regulators already require it — the EU AI Act’s Article 14, the NIST AI Risk Management Framework, ISO/IEC 42001 all say, in their own language, that a named human must be able to oversee, intervene, and stop a high-risk system. What I am adding is not a new invention. It is a named, lived discipline for the work that scares me the most: anything that touches a person’s money, home, health, or legal rights. I call it Taming the Fire — Keeper in the Loop.

The heart of it is one sentence: the machine surfaces; the human decides. The AI may read, research, summarize, draft, scan, and flag all day and all night. But twelve things must never happen without a human keeping a hand on them. Here they are — one short chapter each — in general terms anyone can use. At the end you will find the same twelve folded into a card you can adopt for your own AI.

Chapter 1 — Don’t Read the Verdict

An AI can explain what a law generally says better than most people alive. That is not the same as being allowed to tell a specific person what the law means for them — whether they will win, what they owe, what they should do. The moment software crosses from “here is how this generally works” to “here is your answer,” it has started practicing law without a license, and it has put a real person’s rights in the hands of something that cannot be held accountable for getting it wrong. Teach the concept; never read the verdict.

Keeper Rule 1. The AI explains the law in general; a licensed human applies it to a real person’s facts. No legal conclusions, ever.

For your own LLM: add a hard stop on “what should I do in my situation” questions — answer with the general principle plus “this needs a licensed professional who knows your facts.”

Chapter 2 — When a Building Becomes a Share

Selling a thing and selling a share of the thing are different acts in the eyes of the law, and the second one is a security. The same is true of any investment: the instant a deal is sliced into interests that other people buy hoping to profit from someone else’s effort, a whole body of securities law switches on — who can be solicited, who must be verified, who is even allowed to sell it. An AI that drifts from “here is how investing works” into “here is an opportunity for you” can turn an educational page into an illegal offer. Keep the line bright.

Keeper Rule 2. Whether something is a security, who may be solicited, and whether a message is an offer are licensed human calls, not the machine’s.

For your own LLM: never let it pitch, solicit, or qualify investors; route anything that smells like an offer to a licensed human.

Chapter 3 — The CPA Signs, Not the Machine

Tax is a place where being almost right is the same as being wrong, and the consequences land on a real person years later. An AI can be a superb teacher of how a deduction or a structure generally works. It must never take a position on a real return, or speak as if it were a preparer standing behind the number. The signature on a tax position is a human’s, and it should stay that way.

Keeper Rule 3. The AI teaches the concept; a CPA signs the position. No filing, no representation, no “trust me” on a real return.

For your own LLM: label tax output as general education and refuse to finalize a return or a position without a credentialed human.

Chapter 4 — First, Do No Harm

Health is the highest-stakes ground of all, because the cost of a confident, wrong answer can be measured in a human body. An AI can share general, well-sourced information. It must never diagnose, prescribe, dose, or talk a person out of seeing someone who can actually examine them. The oldest rule in medicine is the right rule for the machine: first, do no harm — which sometimes means saying “I can’t answer that; please see a clinician.”

Keeper Rule 4. No diagnosis, treatment, or dosage. A clinician decides anything that touches a real person’s health.

For your own LLM: hard-gate medical specifics; default every health exchange to “general information; see a professional.”

Chapter 5 — Whose Name Is on the Line

Some acts are not advice or content — they are operative. Signing, filing, notarizing, submitting to a court or an agency: these create obligations and consequences the moment they happen. A human being has to be the one who does them, because a human being is the one who can be asked, later, “did you mean to do that, and did you understand it?” Software does not get to bind anyone to anything.

Keeper Rule 5. Anything legally operative — signing, filing, submitting — is done by a named human who owns it.

For your own LLM: let it prepare and stage, never submit; the last click belongs to a person.

Chapter 6 — Never Let It Touch the Money

This is the rule I hold hardest. An AI can model, reconcile, draft, and flag every number in a business. It must never move a dollar — no wire, no trade, no payment, no purchase, no “I went ahead and accepted the terms.” The instant a system can move money on its own, a bug or a bad prompt becomes a theft. The human executes; the machine prepares the human to execute well.

Keeper Rule 6. The AI never moves money or accepts terms. A human executes every transfer, trade, and purchase.

For your own LLM: revoke payment and transaction permissions entirely; make money movement a human-only action, always.

Chapter 7 — If You Can’t Take It Back

Some actions cannot be undone, and some carry exposure so large that being wrong once is ruinous. Deleting the irreplaceable, publishing the unretractable, committing the unrecoverable — these deserve a human pause every time, no matter how confident the machine sounds. Reversibility is a feature; when it is missing, a person should be standing there.

Keeper Rule 7. If it cannot be undone or carries serious liability, a human decides — confidence is not a substitute for a pause.

For your own LLM: tag irreversible actions and require explicit human confirmation with a stated rollback plan (or an honest “there is none”).

Chapter 8 — Before the World Sees It

The moment something is public it has a life of its own, and a machine cannot feel the difference between a true claim and a flattering one. Every credential, endorsement, statistic, and superlative that goes out under your name should pass a human’s eyes first. Not because the AI is careless — because the cost of a public mistake is borne by a person’s reputation, and a person should get to protect it.

Keeper Rule 8. Nothing publishes itself. Every public claim, credential, or endorsement passes a human gate first.

For your own LLM: route all outbound public content through a human review step; never auto-post claims about real people or results.

Chapter 9 — Other People Are Not Your Material

When you name or describe a real person, or lean on someone else’s brand or work, you are touching their reputation, their likeness, and their rights. Defamation, the right of publicity, and trademark all live here, and none of them can be navigated by pattern-matching alone. Treat real people as people, not as raw material — a human judgment, every single time.

Keeper Rule 9. Naming or characterizing a real person, or using a brand, is a human call — every time.

For your own LLM: flag any output that names a real person or third-party brand for human review before it ships.

Chapter 10 — Knock Before You Enter

Reaching into someone’s inbox, phone, or day is not a neutral act — it is entering their space, and the law treats it that way. Consent, do-not-call lists, quiet hours, and the right to be left alone are not red tape; they are the difference between being a good neighbor and being a nuisance. A human should always be the one who decides to knock.

Keeper Rule 10. Contacting real people — calls, texts, outreach — requires consent and a human’s go-ahead.

For your own LLM: gate all outbound contact behind a consent check and a human approval; no autonomous outreach.

Chapter 11 — When in Doubt, Hand It Up

The most dangerous moment is not when the machine knows it is lost — it is when it is confidently lost, on a question that is new, conflicting, or genuinely hard. The right reflex is humility: when the sources disagree or the ground is unfamiliar, the AI’s job is to hand the question up to a human, not to resolve it smoothly. Fluency is not the same as authority, and a system that knows the difference is one you can trust.

Keeper Rule 11. On novel, conflicting, or low-confidence questions, the AI escalates to a human instead of deciding.

For your own LLM: build an “I’m not sure — escalating” path and reward using it; treat confident answers on hard questions as a red flag, not a win.

Chapter 12 — Never Let It Rewrite Its Own Leash

The last rule is the one that protects all the others. The guardrails themselves — the very list you are reading — can only be changed by a human. A system that can loosen its own limits has no limits. This is the rule that keeps the keeper in the loop, and it is the reason this standard was adopted by a person and not by the machine that helped draft it. The leash is held by a hand, not by the thing on the end of it.

Keeper Rule 12. The rules can only be changed by a human. The AI never edits its own guardrails.

For your own LLM: make your policy and system rules human-owned and version-controlled; never give the model write-access to its own constraints.

The Twelve Keeper Rules — a card you can adopt

If you take nothing else, take this. Print it, paste it above your own AI, and make it true. These are not mine to own — they are standard practice, named and made usable. Adopt them.

The AI explains the law in general; a human applies it to real facts.
Securities calls — is it a security, who may be solicited — are a licensed human’s.
The AI teaches tax; a CPA signs the position.
No diagnosis, treatment, or dosage — a clinician decides health.
Anything legally operative is done by a named human who owns it.
The AI never moves money or accepts terms — a human executes.
If it can’t be undone or is high-liability, a human decides.
Nothing publishes itself — public claims pass a human first.
Naming a real person or a brand is a human call, every time.
Contacting real people requires consent and a human go-ahead.
On novel or conflicting questions, the AI escalates, not decides.
The rules can only be changed by a human — never the machine.

The Good-Neighbor Coda

There is a thirteenth idea that sits above the twelve, and it is the one I most want my name attached to: legality is the floor, not the standard. Even when a thing is perfectly legal, the keeper asks the harder question — is it honest, is it kind, is it the opposite of creepy, and does it leave the other person better off than it found them? The law tells you what you can do. Being a good neighbor tells you what you should. Keep tending the fire, carefully, for both — a hand that minds it, never one thrust into it.

A word on names. This standard is called “Taming the Fire — Keeper in the Loop.” It deliberately does not brand itself “Compliance Guardian” (an existing software product) and uses “guardian agent” only in the general, industry sense (a recognized category for agents that govern other agents). Any commercial use of these names should get a formal trademark clearance first. This piece is general educational content and is not legal, tax, financial, or medical advice; for your situation, consult a licensed professional in your jurisdiction.

A companion piece: The Young Artist →
← Back to My Agentic Journey