Who Holds the Pen? is about agencies putting AI between a caseworker's desk and a decision about someone's benefits. It proposes a judgment router, a layer that sits between what an agent proposes and what actually happens, and decides whether the action goes ahead, gets checked first, or waits for someone with the authority to sign off. The schemas are in an open repository, judgment-routing.
That post is about institutions, where the person holding the pen is a supervisor or a director with signing authority. The same question now lands on anyone who hands an errand to an agent, whether that's booking travel, disputing a charge or moving a meeting, and in those cases the person holding the pen is you.
Four tests
Every action an agent proposes can be scored on the same four signals from Who Holds the Pen.
- Uncertainty. The agent is unsure what you meant, or the facts it found conflict with each other.
- Stakes. Money, a deadline, or something that's hard to undo once it's done.
- Authority. The action goes past what you granted, like a purchase over your cap or a contract term longer than you allowed.
- Novelty. It's the first time this agent has dealt with this company or this kind of task.
Four paths
The scores decide which way the action goes.
| When the action is | What happens |
|---|---|
| Low on all four | It goes through and gets written down |
| Uncertain | It gets checked first, by reading the account back or getting the company to confirm |
| High stakes or past your limits | It waits for your yes |
| Outside what an agent should handle | It goes to a person who knows the domain, like your accountant or a human at the company |
In the agency version these were the fast path, the slow path, human escalation and specialist referral. What carries over is that the paths are written down ahead of time, where you can read them.
What you set ahead of time
In the agency version, a manager writes a strategic context document that sets priorities and authority limits for a whole team of agents. For a person, the same thing comes down to a few plain limits:
- Up to $100 on travel this month
- Access to your calendar for the next 45 minutes
- Negotiate the phone bill and keep the contract at 12 months or less
Each limit has an end date and a way to take it back. An agent carries those limits with it into every task, and the router checks each proposed action against them before anything runs.
The receipt
Every action leaves a receipt that ties what the agent did to the authority that allowed it: what was proposed, which limit it was checked against, which path it took, who said yes, and what happened.
{
"receipt": "rcpt-0142",
"requested_by": "you",
"agent": "travel assistant",
"proposed": "book a $240 train ticket to Seattle",
"limit_checked": "travel, up to $100 without asking",
"path": "waited for your yes",
"approved_at": "2026-01-15T09:12:00-08:00",
"outcome": "booked, confirmation received from the rail company"
}
A receipt like this is what lets you look back a month later and see why something happened, the same way a signed form let an agency reconstruct who approved what.
Why the rules sit with the person
Models read the same instruction differently, and the same model can answer differently depending on how it's asked. So the decision about authority lives in plain rules that the person sets and can read, and the agent's job is to explain what it wants to do and report what happened. That keeps your limits the same across every agent you use, whichever company made it.
Asking at the right moments
The worry with any of this is a phone full of approval prompts that people stop reading. The four tests are how you avoid that. The agent asks where the consequences justify it, and everything else stays in the record for you to look at whenever you want.
Sources
- Ron Bronson, Who Holds the Pen? Why AI Agents Need a Manager, Not Just a Manual, January 15, 2026
- quarterback/judgment-routing, schemas for the router, the four signals and decision receipts