Ben Coumes is a payments engineer whose work spans payment orchestration and the authorization of autonomous commerce: how can a merchant know that an AI agent has permission to spend someone’s money? His background includes a staff software engineering role on PayPal’s enterprise payments team, collaboration on Braintree’s payment infrastructure, and a proposed approach to making consequential agent actions independently verifiable.
Coordinating payments across providers
Coumes collaborated with Sam Parsons on PayPal Braintree payment orchestration. Both are named inventors on a unified API patent application describing how a common interface can coordinate payment operations across providers with different interfaces and notification behavior.
The disclosed design accepts standardized requests and uses provider-specific connectors to translate them into the form each provider requires. It also handles asynchronous notifications and updates a system of record as information arrives. That second responsibility matters: submitting a request and receiving its eventual outcome are separate events, and the orchestration system needs to keep its recorded state aligned with those later updates. The application describes an engineering design; it does not, by itself, establish that the design was deployed in production.
Coumes’s work on agent authorization addresses another part of payment reliability: establishing the permission behind a transaction. Working alongside Jay Mok, he develops a framework organized around three questions: whether the human approved an action, whether it falls within the permitted scope, and whether that permission can be proved later. Their shared framework distinguishes actions by their consequences and by the relationship between the participants. Reversible coding changes may be adequately controlled through tool permissions and recorded in system logs. Payments introduce financial consequences; transactions between unfamiliar parties introduce the additional problem of proving permission without depending on a shared intermediary.
Authorization beyond a shared platform
Coumes advocates independently verifiable authorization for autonomous payments between unfamiliar parties. Within a shared payment platform, buyers and sellers can rely on the platform to enforce the user’s instructions. An autonomous agent dealing with an unfamiliar merchant needs permission that the merchant can verify for itself. He favors convergence on FIDO verifiable intents and AP2 mandates as a direction for addressing that problem.
His layered authorization model separates the responsibilities of the credential provider, the human, and the agent. A trusted provider creates the first layer of a selective-disclosure JSON Web Token. The user signs a second layer containing instructions; autonomous execution can add a third layer signed by the agent. Each participant checks the information relevant to its responsibility: a merchant verifies the checkout authorization, while a payment processor verifies the payment mandate. The proposed mechanism allows those checks without requiring every participant to have an existing relationship with every other participant.
Consent before the purchase is selected
Coumes also describes PayPal’s approval token, which moves consent earlier in the purchasing process. Conventional checkout starts with an item and ends with the user approving its payment. In the agent flow, the user first confirms the agent’s instructions through PayPal, before the agent has found the item to buy. The returned authorization includes constraints such as an amount, an expiration, and a permitted merchant.
That arrangement gives the agent room to carry out a task while preserving limits the human has already approved. It also separates two mechanisms that can otherwise sound interchangeable: the PayPal token is opaque and depends on PayPal for validation, whereas independently verifiable authorization is intended to let unfamiliar participants check permission themselves. At the time of the recording, the approval token was approaching production release; the broader model for autonomous payments between unfamiliar parties remained a proposed direction.
Coumes extends this reasoning to actions that are hard to reverse, including medical orders, electronic signatures, and securities trading. His concern is the permission behind an action: whether the agent stayed within the human’s instructions, and whether another party can establish that before accepting the result. Payments provide the concrete engineering setting for his wider argument about making consequential automation accountable.
Jay Mok and Ben Coumes of PayPal build an authorization ladder from reversible coding work to autonomous payments, showing how consent, scope, and evidence must strengthen as actions become harder to undo and counterparties become less familiar.
Agent authorization must answer three separate questions: whether the human consented, whether the action remains inside its current scope, and whether that authorization can be proved later.
Use lightweight permissions and ordinary logs when actions are reversible and participants share a trusted boundary; stronger proof should earn its complexity by addressing higher consequences or unfamiliar counterparties.
A shared vault plus OAuth scopes can support machine payments inside a known ecosystem because the common operator enforces mandates and retains transaction history.
For autonomous transactions among unfamiliar parties, layered signatures and selective disclosure separate credentialing, human instructions, agent action, checkout verification, and payment verification.
PayPal’s approval-token example reverses the usual order: the human approves constrained instructions first, then the agent searches and transacts within the amount, expiry, and merchant limits.
The ladder extends beyond payments to any consequential, hard-to-reverse agent action, including medical orders, electronic signatures, and securities trading.
