Engineering guide

Agentic commerce failure modes worth testing before launch

The hard part is rarely exposing a product API. It is preserving commercial truth and user control when the catalog, cart, identity, approval and payment state change independently.

Use this as a test plan: each failure mode below describes a condition that should be reproducible in a sandbox or controlled merchant environment. A protocol announcement alone does not cover it.

1. Catalog data goes stale during a decision

An agent quotes a variant, price or delivery promise that changes before handoff. Return current availability, currency, tax and final constraints at the point the buyer acts, then explain material differences instead of silently replacing the offer.

2. A cart mutates twice after a retry

Network retries are normal. Cart, order and payment mutations need idempotency keys, bounded expiry and a record of the prior outcome. Test the same request arriving twice and arriving after the original response was lost.

3. Buyer context leaks into discovery

Discovery may not need a name, email, payment credential or loyalty balance. Use separate permissions for public product retrieval, account linkage and order operations. Test what a read-only agent can retrieve when it has no authenticated buyer.

4. Consent is inferred instead of captured

A product search is not permission to spend. Define what requires a user confirmation, what can be pre-authorized and how a merchant validates the resulting proof. Higher value, unusual category or changed terms should trigger the controls the merchant has documented.

5. Agent identity is treated as a user-agent string

A browser identifier does not prove who made a request or who authorized it. Where a trust protocol is in scope, verify the documented cryptographic fields, expiry and replay protection. Otherwise state the limitation rather than calling automation trusted.

6. Product selection and checkout disagree

The selected variant, price, locale and quantity must travel to the checkout handoff. Test regional availability, promotion expiry, product substitutions and address changes. A buyer must see any material change before they commit.

7. Order status is assumed, not reconciled

A payment acceptance, an order creation and a fulfillment confirmation are different events. Define the terminal states, compensation path and support ownership for each. Test processor timeouts, declines after authorization and merchant rejection.

8. Policies are available only after an irreversible step

Returns, cancellation windows, shipping restrictions and subscriptions affect the decision. Surface them before the buyer confirms. Test whether the agent can obtain a plain-language policy for the exact product and destination.

9. Observability ends at conversion

Store a privacy-respecting trace that ties discovery, handoff, authorization and order outcome together. Without it, a merchant cannot investigate false declines, duplicate orders or an agent selecting an unsupported option.