Every agentic payment system answers “may this agent pay?”
None answer the question that causes the dispute: “why this purchase?”
REASON-GATED PAYMENT REQUESTS
A better precondition for agentic commerce.
Make agents
show their work.
Before an agent moves money, x428 asks it to explain what the user asked for, what it chose, and why. Then it signs the answer and saves the evidence.
See the handshakeSCROLL TO EXPLORE↓01 / 04
01 — THE GAP
02 — THE IDEA
A middleware
first. A protocol
in waiting.
x428 is a drop-in checkout guard, not a new payment rail. It gives a merchant a structured, signed record of an agent’s reasoning before payment proceeds—on cards, ACP, UPI, or x402.
Does not replace AP2Does not settle fundsDoes create evidence
03 — THE HANDSHAKE
An ordinary HTTP precondition, repurposed for an extraordinary problem.
Agent
wants to buy.
It submits the cart to the merchant’s checkout.
POST /checkout428→
Merchant asks:
Show your reason.
A 428 response defines exactly what justification this purchase needs.
428 PRECONDITION REQUIREDJWS→
Agent returns a
signed Reason Object.
Instruction summary. Selection rationale. Mandate. No private chat transcript.
X-REASON: <signed JWS>✓→
Policy checks.
Payment proceeds.
Every verdict enters an append-only, tamper-evident evidence chain.
200 / 402 / 403THE RECEIPT FOR “WHY”
A reason is not a free-text excuse. It is a signed, structured attestation bound to the cart.
REASON OBJECT / v0.1VERIFIED ✓
{
"instruction": "Black high-tops, UK 8, under ₹5,000",
"selection_rationale": "Only in-stock match at this merchant",
"explicitness": "explicit",
"confidence": 0.94,
"mandate_ref": "mnd_...",
"instruction_hash": "sha256:..."
}Verbatim user conversations never leave the agent platform.04 — WHAT CHANGES
Evidence before
money moves.
Set policies for inference, recurring commitments, and spend caps. Export a dispute-ready evidence bundle instead of reconstructing a story weeks later.
Delegation you
can inspect.
A clear record of why an agent bought something—plus a legible trust signal that the merchant rejects unjustified purchases.
Failures with
a way forward.
Specific, remediable rejections let agents fix a missing disclosure or request confirmation instead of failing in the dark.
FIELD NOTES / 001
Permission is not
reason.
A detailed introduction to x428: why agentic commerce needs an evidence layer, what a Reason Object preserves, and how the 428 handshake fits alongside x402, ACP, AP2, and UPI.
Read the briefFIELD NOTES / 001Why, what, and how:
Why, what, and how:
an evidence layer for
agentic commerce.
Why now
Agentic checkout has solved a real problem: a user can delegate a purchase and an agent can present a valid payment credential. But permission alone cannot explain the transaction that follows. When the item is wrong, a subscription was not surfaced, or an agent stretches a standing mandate too far, merchants and users are left with the same missing artifact: a contemporaneous explanation.
That missing explanation is especially costly where settlement is final. On x402 rails, there is no chargeback process to absorb ambiguity. On card rails, merchants inherit disputes without the clickstream or device trail that traditional ecommerce relies on. The result is friction after the fact: refunds, returns, disputes, and a nagging sense that delegation is a black box.
What x428 changes
x428 adds a small, explicit checkpoint before checkout. A merchant returns HTTP 428 Precondition Required with the justification it needs. The agent responds with a signed Reason Object: a summary of the user’s instruction, the selection rationale, a mandate reference where applicable, and the relevant transaction facts. The merchant validates the object against policy, then accepts or rejects it with a useful code.
This is not a claim that software can prove an agent is telling the truth. It cannot. x428 verifies structure, signature, freshness, cart consistency, and policy—not the inner truth of a reason. Its value is disciplined friction and durable evidence: an agent has to commit to an explanation before it spends.
How it stays practical
It is designed as middleware first. An engineer can place an x428 guard around an existing checkout route; settlement keeps using the rail already in place. Returning agents can attach their Reason Object proactively, avoiding the extra round trip. For x402, the accepted evidence reference binds reason to settlement, producing a chain from instruction hash to signed reason to payment and, if necessary, refund.
Privacy is a first-class constraint. The merchant receives a relevant summary and a hash of the original instruction, never the verbatim conversation. The platform retains the full instruction and can disclose it only with the user’s consent if a dispute genuinely requires it. That is what makes the evidence useful without turning every checkout into surveillance.
The ambition is modest and specific: give merchants a new policy lever, give people a receipt for an agent’s decision, and make agentic commerce easier to trust one well-explained purchase at a time.
BUILDING IN THE OPEN