GapLimit

R41Future Friction

Before an AI agent can pay

Plausible
Payment is the last click in a much longer argument. Before software spends one dollar, someone must prove whose intent it carries, what it may buy, how much it may lose, who absorbs fraud, and which transaction can be reversed. The rail arrives after the authority.

The short answer

AI agents may use stablecoins for some machine-readable, cross-border or continuously available settlement. They can also use tokenized card credentials, bank APIs or closed platform balances. The winning rail will be the one that fits authorization, merchant acceptance, liability, identity, compliance and recovery—not merely the one an agent can technically sign.

Original GapLimit object

The machine-payment authorization stack

Settlement is one layer. Remove any earlier layer and the transaction may be technically valid but commercially unusable.
  1. 01
    Human intentWhat outcome was authorized?
  2. 02
    Agent identityCan the merchant recognize it?
  3. 03
    Policy envelopeAmount, merchant, time and category limits
  4. 04
    Payment credentialCard, bank, stablecoin or closed balance
  5. 05
    Liability + recordsRefund, tax, audit and dispute
01

Recognize the actor

Visa's Trusted Agent Protocol is instructive because it begins with recognition: an agent signature, a linked consumer or device identity and payment information. The design uses existing web and payment infrastructure. It demonstrates that agent commerce can develop through incumbent rails rather than waiting for autonomous wallets.

02

Constrain the authority

A useful mandate may specify merchant, category, time, amount, geography, frequency and approval threshold. A private key without policy is broad power, not good delegation. Stablecoin smart contracts can express some controls; account and card systems can express controls too.

03

Assign the failure

When an agent misunderstands a prompt, buys from a malicious merchant or exposes a credential, the settlement rail does not decide liability by itself. Consumer law, network rules, contracts and insurance determine who refunds whom. Irreversible settlement can reduce chargeback risk for a seller while increasing recovery risk for the buyer.

04

Keep accounts after the transaction

A business still needs invoices, beneficial-owner identity, sanctions screening, tax treatment, expense classification and audit evidence. A machine-readable payment does not make these obligations disappear. The most valuable agent rail may be the one that produces the cleanest evidence package.

Thesis audit

Pressure the bridge

Causal bridge
Autonomous task → delegated commercial intent → verifiable agent and policy → accepted credential → settlement → accounting and dispute.
Counterforce
Card networks, bank APIs and platform balances already bundle identity, acceptance and dispute rules; stablecoins must outperform that whole bundle for the use case.
What would prove it wrong?
The stablecoin-first thesis weakens if most agent transactions remain domestic, reversible and embedded in merchants that already accept tokenized card or account credentials.
Largest uncertainty
Agent commerce standards and products are still changing. This is an architecture comparison, not a forecast of transaction share.

Source ledger

Evidence carrying this piece

Sources accessed 2026-09-06. Links point to the originating institution where available.
  1. Visa · Trusted Agent Protocol specificationsCurrent agent recognition, identity and payment model
  2. NIST · AI Risk Management FrameworkAccountability, risk and lifecycle framework
  3. Circle · USDC documentationProvider documentation for a programmable stablecoin settlement rail