What the Approver Sees Before the Tap: Evidence, Door and Blast Radius
The first objection to a person’s tap before every agent write is always the same one: the approver will either rubber-stamp or become the bottleneck. It is a fair objection, and both endings happen. They happen for one reason. Every request looks the same, so the approver has to spend the same attention on a CRM field as on an email to a customer, and spread evenly, attention runs thin.
The answer is not fewer taps. It is a request that says how much attention it needs, and carries what that attention has to look at.
Why approvers rubber-stamp
A typical card reads: “Updated account status to churn risk.” The approver cannot check that from the card. Checking means opening the CRM, the ticket the agent read and the account history, three tools for one tap. Either the approver opens them, and the queue becomes the bottleneck, or the approver taps, and the tap is a stamp.
The card for an email to a customer usually looks the same: one line of summary, one button. The approver has no way to tell from the card which of the two requests could be put back in a minute and which could not.
Three things every request carries
Matt Pocock’s /pr skill for coding agents builds a pull request body the same way, to make human review as easy as possible: a summary, then hard evidence, then the merge danger, framed as a one-way or a two-way door, and the blast radius. The same three fields carry over from pull requests to every write an agent proposes in tickets, chat, wiki and CRM.
Evidence the agent produced, not a claim it made. Before and after on the real record: the current field value and the new one side by side, the passage from the ticket the proposal was built from, the customer it resolved and the field that matched. For an email, the email exactly as it will go, with every recipient. Without evidence, an agent will say the change probably works because it read the code, or the ticket, and the approver is left to believe it.
The door. A two-way door can be walked back through: a person puts it back from inside your own systems in one write, because the record carries its previous value. A one-way door does something out in the world, deletes data, touches money or is expensive to roll back. That is the same split as reversibility without a rollback layer, now printed on the card.
The blast radius. How much the action touches, as a count: one field on one account, a status on 400 tickets, an email to a distribution list of 200 people, an amount in euros. A two-way door with a blast radius of one is a different request from a two-way door with a blast radius of 400, even though both can be undone.
The review depth follows from the door and the blast radius
| Request | What the approver looks at | How long it takes |
|---|---|---|
| Two-way door, small blast radius One CRM field, one ticket status | The before and after on the card. Nothing has to be opened. | Seconds. Once the audit trail shows the class is clean, it can move to review after the fact. |
| Two-way door, large blast radius A bulk status change | The count, the filter that selected the records, and three samples with before and after. The undo is one batch. | Under a minute. |
| One-way door, small blast radius One customer email, one order | The exact content as it will go out, the recipient, and the passage it answers. | A real minute of reading. |
| One-way door, large blast radius A mailing, a payment run, a deletion | All of the above, and a second person, or the action split into smaller requests. | As long as it needs. Never a single tap from a phone. |
Most agent writes are small two-way doors, so most taps are the seconds kind. The attention saved there is what pays for the minute on the few one-way doors. That is how one person can approve every write without becoming the bottleneck, and without the tap turning into a stamp.
Two rules keep the table honest. A request that arrives without evidence is marked as such and moves one row down, never up. And the depth is set by the action class, so a request cannot ask for less attention than its class demands.
The agent may not grade its own request
If the agent fills in the door and the blast radius, the card is a paraphrase again, only with more fields. An agent that labels a bulk email “two-way, small” has not lied on purpose; it summarised what it intended, and the summary is wrong.
So the door comes from the action class, defined once per tool and system, not per request. The blast radius is counted by the harness from the arguments that will run: the recipient addresses, the records the filter matches, the amount. The before value is read from the system of record, not reported by the agent, and a dry run, where the target system has one, is executed by the gate. The agent proposes the write and points at its sources. It does not decide how closely it gets looked at.
This is the same rule as for code: a signature is not evidence when the evidence comes from the system that did the work. An approval is only as good as the card it was given on, and the card is only as good as the part of it the agent could not edit.
What the card holds
On every request, before the tap:
- the action, in the arguments that will run, not the agent’s summary of them;
- the evidence: before and after on the real record, the source passage, the dry-run result where there is one;
- the door, with the action class it comes from;
- the blast radius, as a count taken from the arguments;
- the review depth that follows, and who else has to approve when it is more than one person;
- what happens on decline, and whether anything downstream has already run.
After the tap, the same card becomes the receipt: the six fields of the write that happened, with the evidence the approver saw next to it.
What changes for the team
The approver stops opening three tools per tap, because the card carries what the tap is about. Small reversible changes take seconds, and the one-way doors get the minute they deserve. And the security review gets a rule it can test on the trail: no one-way door was approved on a card without evidence, and the blast radius on every receipt matches the write that ran.
The self-check has the ten-minute tests for your own approval cards, starting with the one that matters here: does the card show the arguments that run, or the agent’s summary of them?