Rails can bridge buyer instruments. A wallet or card token can resolve the same 402 through an interoperable router; chat runtimes rarely block a third-party rail and platform take stays near card interchange. By 2029Q2 24% of agent attempts succeed, agents carry 22% of online GMV — inside Bain's 2030 band a year early — and a third of buyers hold a wallet. The binding constraint flips from the buyer's instrument to seller reach.
The buyer still states the same intent and grants the same bounded authority. What changes is that identity, payment and receipts can cross runtimes, so agents compete on judgment instead of captivity.
One constant: the selected care coordination need and its human boundary. One variable: who can interpret that authority and where recourse lives.
By 2029 the buyer-side instrument is no longer the main constraint. The story can move to seller agents, negotiated capacity and policies that remain valid across providers.
illustrative emerging mechanism · coordinated open rails · not the default 2029 purchase · illustrative, not a forecast
The near-term engine’s settled agent GMV becomes the initial delegated-demand pool. The structural layer then separates demand an agent can influence from the smaller share it can execute under authority, seller coverage and governance.
The major chat runtimes keep third-party rails enabled. A card token and a wallet resolve the same 402 challenge through an interoperable router, so an agent tries whichever the buyer has and falls through to the next. Success reaches 11% and 47% of sellers answer at least one rail.811
The big storefront platforms ship native 402 endpoints as an update, and x402 and MPP begin clearing orders at mid-market merchants that never wrote a line of payment code. Agents route around page-only sellers, suggesting a checkout-capable equivalent mid-task — a nudge merchants feel as lost orders.37
Success passes 14% and agent GMV share reaches 3.5%. With compatible rails on more checkout-capable sites, agents compare total cost across rails in parallel; merchants offering more than one see the higher modeled conversion. Onboarding flows pre-link an instrument when a person installs an assistant.7
Page-only sellers now serve 42% of modeled demand, down from 60% at the start, while 56% answer a rail. Cross-border orders become easier to represent because the 402 handshake can carry currency and tax terms, and x402 and MPP each hold just over a fifth of settled volume.37
Nearly six in ten sellers expose a rail. Replenishment and subscription purchases — detergent, groceries, renewals — make up an outsized share of agent GMV, because low-stakes, high-frequency orders tolerate the remaining friction. 'No rail' moves beside 'no instrument' as the market's next binding failure.7
Agent GMV share passes 10%. Some sellers answer 402 with terms, not just a price — bulk discounts, slower shipping for a lower total — and a few publish agent-only bundles to see whether algorithmic buyers shift demand. This is where the scenario begins turning checkout into negotiation.11
'No rail' overtakes 'no instrument' as the top failure. Buyers are ready; the long tail of sellers is not. Agents fail over card → x402 → MPP routinely and learn which merchants settle fastest; people stop setting a default payment method and let the agent pick per order.79
Agents carry 22% of online GMV with a 24% success rate, and 34% of buyers hold a wallet. The rail mix is five-way — x402 22%, MPP 21%, UCP 20%, card tokens 18% and platform checkout 17%. Interoperability, not standardisation, is what won; agents compete on routing, not exclusivity.678
Continue this institution into the 2040 story →
The market develops coordination capacity that bridges instruments, lowers seller integration effort and exposes failed demand. The writer receives this cross-quarter spine, all actor arcs, counterfactual outcomes and open questions before drafting any scene.
Across unseen stores, how often does the agent finish a purchase and recover from a changed price or unavailable item? Evidence checked 2026-09-04: Anthropic publishes shopper and merchant agent harnesses with tools, memory and approval boundaries. Limit: A working architecture does not establish independent completion rates, merchant coverage or error recovery. Test: Repeat a fixed task set across merchants; record successful outcomes, human interventions and recovery, including failed attempts. Which merchants support checkout, fulfillment changes and refunds end to end with a delegated mandate? Evidence checked 2026-09-04: UCP defines a checkout lifecycle; its base flow requires human finalization unless the AP2 Mandates extension is supported. Limit: A published checkout interface does not demonstrate broad deployment or unattended completion. Test: Test purchase, cancellation and refund across independently implemented merchants; distinguish supported endpoints from successful transactions. Can the same agent complete a shopping task across unfamiliar stores without losing fit, total-price or delivery constraints? Evidence checked 2026-09-04: Computer-use tools expose screenshots, mouse and keyboard actions; commerce protocols separately expose structured merchant operations. Limit: Tool availability is not evidence of dependable purchasing across stores, changing interfaces or exceptions. Permission alone does not supply that ability. Test: Hold the shopping brief fixed across unseen stores; record completed orders, missed constraints, interventions and recovery after UI changes. Can a mandate be revoked or constrained across two independent agents, merchants and payment implementations? Evidence checked 2026-09-04: AP2 describes human-present and human-not-present flows using signed mandates. Limit: Signed authority alone does not establish interoperable revocation, dispute outcomes or fidelity to a person’s intent. Test: Exercise expiry, revocation, budget changes and replay across implementations; verify where control returns to the person. What observed purchase-attempt series would justify replacing exogenous intent growth with an estimated adoption process? Would merchants actually respond to failed agent demand quickly enough to reach 65.0% rail coverage? Which runtime policy, distribution or regulatory events would change third-party rail access? How much cross-instrument reach and seller integration savings can ecosystem-wide coordination cause beyond open protocols alone, and which institutions could supply it? Which reproducible commercial-task evaluation series could anchor capability learning after 2029 instead of treating it as a structural prior? Will merchant-native commerce-agent blueprints produce durable shopper relationships and governed pricing or inventory delegation, rather than remaining ordinary chat and merchant automation? Which standards, statutes and dispute outcomes would demonstrate that authority, receipts and liability are becoming portable across runtimes? What comparable loss, dispute, collusion and manipulation datasets could anchor adversarial pressure as machine counterparties grow?
Each lane is generated from the same quarterly state as the charts. The action is a deterministic interpretation of pressure in the model, not prose invented after the fact.
By 2029Q2: 25% of attempts settle, agents carry 22% of online GMV and merchants retain 22% after platform take; 35% of buyers hold a wallet and 65% of sellers answer a rail. Top failure: no rail at 25%.78