Before you build an AI agent, write the stop rule

The most important line in an AI agent spec is the one that says when the agent stops and hands the work to a person. Most specs leave it blank. The blank gets filled during an incident, by whoever is awake.

Write the stop rule first. Before the model choice, before the integration plan, before the demo. It fits on one page: three questions, three names, three clocks.

The stop rule belongs to the operator

An agent with no stop rule takes the decisions it should have asked about and leaves people the approvals, the exceptions and the failures. The cost shows up as a refund a customer should not have received, a vendor paid without a PO, a reply to a legal threat sent in a support tone.

The stop rule is where an operator decides which decisions the company keeps. That is a business call, and it belongs in the spec first.

Three questions the agent asks before it acts

Write them in prose; the flowchart will not be open during an incident.

1. Does this action move money, send anything outside the company, or change what a customer sees? If yes, stop and wake the named approver.
2. Does this case match a condition on the expertise list? An unfamiliar vendor, an amount above the threshold, an account flagged by its manager. If yes, stop and route to the named specialist with the case attached.
3. Is this a decision the team said it wanted to keep? If yes, prepare the options and stop.

If the answer to all three is no, the agent acts, logs the action in plain language, and moves to the next item.

Three names

Each question has a person on the other end of it. Write the names. A role title is where escalations go to die.

• The approver for question one: someone with authority to sign for money and outside contact.
• The specialist for question two: the person whose expertise the model lags, the controller, say, or the senior estimator.
• The decision owner for question three: whoever wants to keep the interesting calls.

Three clocks

Each stop creates a waiting item, and the third thing to write is what happens to it. A wrong action gets noticed within the day. A silent queue gets noticed on Thursday, about Monday's work, by the customer.

Put a time limit on each stop. Wake the approver at zero, the approver's manager at four hours, and the operator who owns the workflow at end of day.

Three short examples

These are illustrative, not client cases.

A quoting agent prices a standard order for a known customer at list, acts and logs. The same agent gets a 30% discount request on an order twice that customer's largest. Two conditions fire, so it stops and wakes the sales lead with the order attached.

An invoice matching agent finds a PO, a receipt and an invoice that agree, and marks the invoice for payment. It finds an invoice from a vendor added last week with no PO. The expertise rule fires, the controller gets it with the vendor record attached, and the agent does nothing else with that invoice until a person clears it.

A support drafting agent answers a shipping question and sends the reply. A message mentions a lawyer. The approval rule fires, the draft goes to a person, and the thread is marked waiting.

Evidence: Mollick's four reasons to hand work back

Ethan Mollick's "Agency and Agents" (One Useful Thing, August 31 2026) makes the same case from the other direction. Rather than asking when a person should hand work to an agent, he asks when the agent should hand work back. His design, the Twilight Factory, has agents doing most of the work and reaching out to people on purpose rather than on failure.

Source facts: he gives four reasons to reach out. Approval, for spending money, contacting outsiders and touching sensitive material. Expertise, because models are jagged and lag human experts on parts of the work. Variance, because AI output repeats the same patterns, themes and even character names. Interesting work, because if agents make each interesting decision and leave people the approvals, exceptions and failures, "we will have automated the wrong half of the job."

My read: approval, expertise and interesting work map onto the three questions. Variance is a quality problem rather than a stop condition, and it belongs in the monthly review sample.

The month-one test

Run the agent for a month and count escalations. A handful of stops a week means the rule fires and the thresholds need tuning. Zero stops means one of two things: the rule does not fire, or the scope is narrow enough that nothing qualified. An agent with a narrow scope can produce zero on purpose. Find out which you have by feeding it one case that should stop it. If it acts, the stop rule was decoration.

Three questions, three names, three clocks, written before the first line of code. The agent will be replaced within a year. The stop rule will outlive it.

If you are speccing an agent this quarter, I can send the one-page stop rule template: the questions, blanks for the names, defaults for the clocks. Reply with the workflow and I will send it.