ARTICLE
Company
The rails for machine-to-machine payments now exist. The layer that decides which of those payments should execute largely does not. That gap — between a rail that can move value and a control that decides whether it should — is the open problem in autonomous commerce.
What has already been built?
Standards for agent-initiated payment arrived quickly. Google announced the Agent Payments Protocol (AP2) on 16 September 2025 with more than 60 partner organisations, using signed Intent and Cart Mandates carried as Verifiable Credentials to record what a user actually authorised.
Alongside it, machine-payable HTTP endpoints and stablecoin settlement now let software pay software with no person present at the moment of payment.
Why aren’t rails enough on their own?
A rail is neutral by design. It moves value between two parties and asks very little about why.
That neutrality is what makes a rail useful, and what makes it insufficient alone: the same properties that let a legitimate agent pay for an API call let a compromised one drain a budget.
What belongs in the control layer?
Authorization is largely addressed by mandates — cryptographic proof that a user granted an agent authority. What sits above the rail is the runtime decision: whether this specific payment, at this moment, for this stated reason, should execute.
Every previous generation of payment infrastructure grew that layer eventually: authorization, fraud checks, spending policy. Autonomous commerce has not grown one yet.
Frequently asked questions
Don’t mandates already solve this?
Mandates establish that a user authorised an agent to act. They do not evaluate whether an individual payment matches the task, fits the remaining budget, or repeats a payment already made.
Is this a blockchain problem?
No. The control layer is independent of the settlement rail. The same decision — allow, deny, hold — applies whether a payment settles in stablecoins, over card networks, or by bank transfer.