Skip to content
← Insights
Nov 15, 20264 min read

An agent's purchase needs an authorization the executor can verify

Bind spending approval to a specific proposal, enforce it outside the model and reconcile uncertain transaction results.

A purchase proposal card passing through a small approval gate toward a payment card, with a lock and receipt.

An agent researching a paid service does not need permission to buy it. The moment it can accept a trial or upgrade a plan, the workflow has acquired a financial capability that needs its own boundary.

I would make that boundary visible in the system design. The component that researches options can propose a transaction. A separate executor checks whether the proposed transaction fits the user's actual authorization.

Put the proposed purchase in front of the approver

"Continue" is an ambiguous approval label when the next step starts a subscription. Show the merchant and the account being changed. State the amount and currency, then distinguish a one-off charge from recurring billing.

A trial needs its conversion terms as well. A zero initial price does not describe the later obligation. If those terms cannot be established, the proposal is incomplete.

Bind approval to the proposal that was reviewed. A change in price or billing period should invalidate an approval whose scope no longer covers the transaction. The agent's summary of an earlier conversation is insufficient evidence for expanding that scope.

The AP2 project's agent authorization model distinguishes user delegation from verification of a particular action. It describes checking both mandate integrity and permitted scope. Treat this as a specified model that an implementation must support, not a feature every merchant already offers.

A signature can protect the integrity of a record. The verifier still needs to evaluate its contents correctly. Signing an overbroad or misunderstood instruction preserves that problem.

Enforce the limit outside the agent's editable state

A prompt saying "spend no more than the approved amount" expresses policy. It does not enforce that policy if the same agent can alter the executor or use another credential.

Give the executor only the capability it requires. Keep approval validation and spending-state updates outside the model's control. Test that a rejected proposal cannot reach the payment operation through another tool.

A cumulative budget needs a concurrency policy. Two individually acceptable requests can exceed a shared total if both check the remaining budget before either reserves its amount. Reserve against the same authoritative ledger before allowing execution.

Payment infrastructure can supply another control layer. Stripe Issuing's spending controls describe limits and restrictions for eligible card programmes. The documentation also notes best-effort aggregation, including possible delay, and circumstances where later charges can exceed a limit.

Those limitations belong in the design. A card limit is useful, but it should not be described as a perfect atomic guarantee for every form of cumulative spending. Verify the capabilities available to your specific programme.

A timed-out purchase needs reconciliation

Suppose a hypothetical executor sends an approved purchase, then loses the connection before receiving the result. The remaining question is whether the merchant accepted it.

Do not treat a missing response as a declined payment. Check the merchant or processor record using the transaction identity. If the outcome remains unknown, preserve that state for reconciliation.

Where the API supports it, reuse the appropriate idempotency identifier for retries of the same operation. Stripe's idempotent-request documentation defines the behavior of that mechanism for its API. Check its retention and parameter rules before relying on it.

Idempotency and permission address different errors. Repeating one unauthorized charge only once still leaves an unauthorized charge. The executor needs both an authority check and duplicate prevention.

Store a receipt that connects the approved proposal to the observed result. Minimise sensitive payment data in that record. An investigator needs enough information to reconstruct the decision without giving every log reader payment access.

Recurring authority needs an end condition

A monthly subscription can continue after the original task is finished. Record whether the user's authorization covers renewal, with the maximum scope and any expiry.

A preference in assistant memory should not silently become permission for a new merchant or a higher tier. The executor should require a valid authorization for the actual proposal.

Test the unhappy paths before granting real spending access. Change the currency after approval. Submit an expired authorization. Send the same request again after an ambiguous response. Confirm each case reaches the intended refusal or reconciliation path.

The desired result is a purchase whose authority can be explained after the conversation is gone. The approval record should tell you what the user permitted, and the transaction record should show what the system actually did.

AI agentsPaymentsSecurity