An AI assistant is shopping for you. Halfway through, a box pops up:
1Acme Shop wants to use create_order2 3 [Allow once] [Always allow] [Deny]
You click Allow once. It feels like control. The agent asked, a human was in the loop, the human said yes.
But what did you decide?
Not much, honestly. You decided the assistant may create an order. You did not see the item, the price or the delivery address. If you asked for a monitor and it is ordering a €749 gaming chair, your yes approved the chair.
A lot of agent approvals look like this, I think. They put a person in the loop and then give that person almost nothing to decide with.
Three questions
When I look at an approval step now, I ask three things:
- What did the person actually see?
- What does "yes" cover?
- Is the "yes" tied to what actually happens?
The first one usually gets some attention. The other two are where it gets interesting.
What did the person see?
The MCP spec has the right instinct. Clients SHOULD "show tool inputs to the user before calling the server". So let's be generous and show the arguments:
1{2 "tool": "create_order",3 "arguments": { "sku": "CH-2231-BLK", "qty": 1, "shipping": "std" }4}
Better. Is that a monitor?
Tool arguments are written for the API, not for a person. The SKU is perfectly precise and completely meaningless to me. What I need to decide on is "a black gaming chair, €749 including delivery, from this supplier, to my home address". The shop knows all of that. The client drawing the prompt usually doesn't.
Then there is the friendly summary. Plenty of prompts add a line like "Order ergonomic upgrade for home office". Readable, helpful, and written by the same model which is asking for permission.
Claude Code gets this layout right. The short description above a shell command comes from the model; the command underneath is what will actually run. The description is a hint, the command is the fact. If an assistant can talk itself into buying a chair, it can write a pleasant one-line summary of the purchase too.
So a useful prompt shows the concrete action in human terms, described by something other than the party asking.
What does "yes" cover?
"Always allow" is where a decision about one order quietly becomes a standing permission.
I click it because it is prompt number eleven and I want lunch. What I meant was "yes, this
monitor is fine". What got saved is "this assistant may call create_order on Acme Shop".
Any item, any amount, from now on.
Claude Code is pretty careful here too. Its prompt spells out the rule it would save
("Yes, and don't ask again for: npm test *"), and the docs say it offers that option
only "when the prompt can show you everything they would allow".
That is the right principle: a person can only delegate what they can see.
Even so, a command prefix is a blunt tool. The same docs warn that patterns which try to constrain command arguments are "fragile". Fair enough, they are matching shell strings.
For purchases, an honest "always" would be something like "monitors under €800, delivered to my home, until Friday". That is not a button anymore. That is a policy. Which is fine, but somebody has to design it, and "Always allow" is not that design.
Is the "yes" tied to what happens?
This is the one I find most neglected.
Say the prompt was great. A specific monitor, €729, delivered Thursday. You approve. The order fails because the item just sold out, the assistant retries with "a similar product", and the second call goes through.
Did you approve that?
If the approval only lives in the client, nobody downstream can tell. The shop's API gets an ordinary request with an ordinary token. It has no idea whether a person approved anything, let alone which version of the order.
Banks had to solve exactly this. Under the EU rules for strong customer authentication, when you confirm a payment, "the payer is made aware of the amount of the payment transaction and of the payee". The authentication code is "specific to the amount of the payment transaction and the payee", and "any change to the amount or the payee results in the invalidation of the authentication code generated". They call it dynamic linking.
That is what I want from an agent approval. Not a click, but evidence: a record bound to the concrete action, which the place doing the purchase can check. Change the item or the price and the approval stops matching.
None of this needs new cryptography. OAuth's Rich Authorization Requests can already carry structured details, like an amount or a product, instead of a vague scope name. The harder part is the flow: the resource describes the action, the person approves that description, and the resource checks the approval before it acts.
Wrapping up
An approval prompt is an authorization decision made by a person, in about two seconds. It deserves the questions we would ask any other authorization decision. What exactly is allowed? For how long? Who checks it, and where?
A yes to create_order is a yes to almost anything. A yes to "this monitor, this price,
this address", bound to the order it approves, is an actual decision.
Next time an agent asks me for permission, I will try to notice which one I am being offered.