four-party-model · agentic-commerce · iso8583

Why the 4-party model breaks for agents

The card network architecture that has moved money for decades assumes a human is present. Agent payments do not replace the rails; they replace the trust signals around them.

The four-party card model is usually drawn as a settlement loop: consumer, merchant, acquirer, issuer, with a network in the middle. For agentic payments, it is more useful to see it as a layered trust system.

Each participant asks a security question:

Those questions were built around a human being present in real time. Agent payments break that assumption at the root.

Defense stackThe 4-party model is a fraud system before it is a settlement diagram

Every layer reads a version of human presence. Agent payments rebuild those reads with cryptographic evidence.

Old root signalhuman present and reactive at transaction time
removed by delegation
New root signalhuman who delegated, with policy and evidence, before transaction time
L3 trust invariantRegistered instruction

The network no longer relies only on behavior. It compares the authorization message with a consumer-authorized instruction registered before checkout.


Layer 1: merchant and gateway

The gateway filters automated traffic before payment. CAPTCHAs, WAF rules, device fingerprints, mouse movement, residential IP reputation, and request velocity all exist to stop card testing and credential stuffing.

A legitimate shopping agent can have the same external shape as hostile automation: cloud origin, programmatic browsing, fast request timing, and no human motion signal. The fix is not a better CAPTCHA. The fix is changing the question from “does this look human?” to “did a registered agent sign this request?”

That is why TAP, Web Bot Auth, and AP2-style request evidence all converge on signed messages. Public-key signatures make the request attributable. Nonces and timestamps make it fresh. Body digests make tampering visible.


Layer 2: acquirer and processor

If the request reaches payment, the acquirer still sees a card-not-present authorization. The problem is that legacy payment messages were not designed to explain that a software agent is acting under prior consumer authorization.

The replacement is not a new rail. It is richer context around the same rail:

Public network documents do not currently give one universal ISO 8583 field map for agentic commerce. Any field-level diagram should be labeled as illustrative unless it comes from a published network specification.


Layer 3: network

The network sees cross-merchant behavior that no single merchant or issuer sees. That is valuable, but it also means machine-speed checkout can resemble the attack patterns those systems were built to catch.

The missing signal is prior intent. A legacy authorization does not, by itself, prove the consumer already authorized this merchant, amount, item, and agent.

Agentic payments add an evidence object before checkout: an instruction, mandate, or token-bound intent. When authorization happens, the payment stack can compare the live transaction against that evidence instead of relying only on behavioral inference.


Layer 4: issuer

The issuer is still the final authorization decision maker. In human checkout, ambiguous risk can trigger 3DS or banking-app step-up. In delegated commerce, live step-up is often the wrong fallback. If every useful agent purchase needs a real-time user challenge, the delegation failed.

The replacement is pre-authorized policy: amount caps, category limits, merchant constraints, velocity, expiration, token status, and revocation. Passkeys and other strong authenticators matter because they move authentication to enrollment or instruction time, where the consumer is actually present.

The issuer’s question changes from “is the cardholder here right now?” to “does this transaction fit a policy the cardholder already authorized?”


The four primitives

Where the VIC architecture post maps Visa’s three building blocks — VIC API, TAP, and DCAP — this post maps the four trust primitives any agent payment system has to rebuild, regardless of network or protocol. They are composable requirements, not a Visa-specific stack.

Cryptographic agent identity. A verifier accepts the request because an accountable agent signed it, not because it resembled a human session.

Agent-bound credential. A payment credential should be scoped to the agent and context that received it. A stolen token without the agent’s key or authorization context should not be sufficient.

Registered instruction or mandate. The payment system needs an object that says what the consumer authorized before checkout.

Pre-authorized spending policy. Issuers, wallets, or networks need deterministic rules that can be checked without waking the user at transaction time.

Those primitives are compositional. Identity without a scoped credential is not enough. A scoped credential without intent evidence is not enough. Intent evidence without issuer policy still collapses into step-up when risk rises.

Primitive mapEvery workable design rebuilds the same four trust signals

Products differ in packaging. The architecture keeps converging on the same replacements.

P1
Merchant + gatewayCryptographic agent identity

The request is signed over method, authority, path, digest, timestamp, and nonce.

replacesCAPTCHA, device fingerprint, IP reputation
evidence on the wire
Signature-InputSignatureagent public key
P2
Acquirer + processorAgent-bound credential

The credential only works when presented by the agent identity it was issued to.

replacesReusable PAN or token without agent context
evidence on the wire
tokenized credentialagent contextnetwork metadata
P3
Card networkRegistered payment instruction

The network compares the authorization message against the consumer-authorized instruction.

replacesML risk score without prior intent proof
evidence on the wire
payment scopemerchant matchamount match
P4
Issuer bankPre-authorized spending policy

The issuer enforces deterministic rules created before the transaction.

replaces3DS challenge to a live cardholder
evidence on the wire
Passkey assertionlimitMCCvelocity
Implementation fitSame primitives, different ownership boundaries
system
P1identity
P2credential
P3instruction
P4policy
architecture note
VisaTAP + VIC + VTS
native
native
native
native

Illustrative ISO 8583 mapping: entry context, additional data, and private-use fields can carry agent evidence downstream.

MastercardAgent Pay + token service
native
native
native
native

Token-first design. More of the binding lives in token metadata and edge verification.

AP2Mandates + payment extensions
partial
paired
native
partial

Strong instruction evidence. Card credential binding depends on the payment method extension.

x402HTTP 402 payment transport
paired
native
partial
paired

Excellent for internet-native settlement. Regulated card-style policy still needs companion identity and authorization layers.


What the rails do not have to change

The easiest mistake is to assume agent payments require new payment rails. They do not.

Card authorizations can still travel through merchant, acquirer, network, and issuer infrastructure. Clearing, settlement, refunds, disputes, and merchant acquiring remain the durable parts of the system. The new work is around proof: what the agent signed, what the consumer authorized, which credential was released, and what evidence remains after the order.

Stablecoin and HTTP-native protocols solve a different problem: internet-native payment and programmable settlement. That can be powerful for API calls and agent-to-agent commerce, but it does not automatically recreate issuer policy, consumer dispute handling, network tokenization, or chargeback evidence.

Composed architectureThe rails stay boring. The trust substrate moves forward in time.

Enrollment and instruction carry the hard trust work. Authorization becomes evidence matching plus policy enforcement.

EnrollmentMove trust work before checkout
01
ConsumerCreates passkey and policy

$500 limit, allowed MCCs, velocity, revocation rules.

P4
02
NetworkIssues agent-bound credential

VCN, Agentic Token, or payment-method extension.

P2
03
AgentRegisters public key

Directory entry lets merchants verify future requests.

P1
InstructionCreate auditable intent
04
Consumer + AgentConstructs payment instruction

Merchant, amount, item, agent, expiration, policy context.

P3P4
05
NetworkRegisters instruction reference

Integrity hash becomes the object later carried in the authorization.

P3
TransactionCarry evidence through existing rails
06
AgentSigns checkout HTTP request

RFC 9421 signature covers method, authority, path, digest, timestamp, nonce.

P1
07
MerchantVerifies agent identity

TAP SDK, Web Bot Auth at edge, or equivalent verification path.

P1
08
AcquirerBuilds ISO 8583 authorization

The standard card rails stay in place. The message fields change.

P2P3
AuthorizationApprove without live step-up
09
NetworkMatches authorization context to instruction

Amount and merchant tampering fails before issuer routing.

P3
10
IssuerRuns deterministic policy checks

Limit, category, velocity, token status, revocation. No 3DS fallback.

P4
authorization payloadMTI 0100 carries the new evidence in old rails
tokenagent-bound credentialcredential
contextagent initiationrouting signal
identityagent evidencecontext
scopeinstruction bindingintegrity

Still unsolved

Multi-agent delegation. If a planning agent delegates the purchase to a subagent, the evidence chain needs to show who held authority, who executed, and which policy applied.

Cross-network identity. A Visa-recognized agent and a Mastercard-recognized agent are not automatically the same principal. Shared directories are technically possible and commercially hard.

Dispute evidence. A chargeback file for an agent workflow needs richer proof than a normal ecommerce session: original instruction, agent request, merchant response, fulfillment, refund, revocation state, and user-visible consent.

PSD2 Strong Customer Authentication. Existing European SCA guidance predates autonomous checkout and does not yet give a clean, agent-specific answer for delegated card transactions.


Sources