Platform tokens are exclusive. Chat runtimes block third-party rails almost half the time, take rates run at a tenth or more, and the open protocols cost a merchant three times as much to integrate as the platform's own checkout. Agents still grow — 15% of online GMV by 2029Q2 — but on cards and platform checkout, with wallets stuck below a fifth of buyers and cross-garden purchases a coin flip.
The buyer can still hand work to an agent, but the identity, instrument and recourse remain attached to a runtime account. Reliability improves inside each garden while authority becomes harder to move.
One constant: the selected ski goggles need and its human boundary. One variable: who can interpret that authority and where recourse lives.
By 2029 the agent can buy, but only inside boundaries chosen by the platform. The long-horizon question becomes whether persistent commerce deepens that convenience or makes exit prohibitively expensive.
illustrative emerging mechanism · walled gardens · 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 largest agent runtimes ship checkouts that only honour their own token. An x402 or MPP challenge from an outside seller is silently dropped in nearly half of attempts. Success dips this quarter even as volume grows, because the successful orders concentrate inside each garden.810
The runtimes push card tokens — lower friction, higher approval — and wallet adoption stalls near 13% as people face incompatible instruments accepted by different merchant coalitions. Platform checkout and card tokens split roughly half of settled volume.5
Merchants integrate whichever proprietary SDK matches where their traffic comes from; x402 and MPP plateau near a tenth of volume each among open-web sellers. 'Runtime blocked' becomes a visible failure class, but missing instruments and missing rails still dominate.8
Success reaches 11% and nearly half of buyers now store a card. The mix stabilizes: platform checkout and card tokens a quarter each, UCP a little more, and the open rails roughly a tenth each.7
Agent GMV share passes 5% and a majority of sellers expose a rail. The volume is real, but agents work inside each garden's silo; a purchase that crosses from one runtime's token to another seller's preferred rail fails as often as it succeeds.7
Processors publish monthly agent-acceptance scorecards for their merchants. New sellers pick one of five fixed options rather than invent a sixth. Page-only sellers now serve 37% of modeled demand because packaged capability is cheap, not because the rails opened.7
Only a third of buyers have nothing linked, yet the top failure is still the instrument — often the wrong one for the garden the seller lives in. Wallets sit at 18%. This is what the scenario produces when the buyer's instrument belongs to the platform.9
Agents carry 15% of online GMV with a 17% success rate — seven points of GMV below the open ending on the same demand. Cards clear 25% of settled volume, platform checkout 24%, and x402 and MPP about 12% each. Some verticals may see agents drive a quarter of orders; nobody sees agents buy anywhere.678
Continue this institution into the 2040 story →
Runtimes privilege proprietary credentials and checkout, fragmenting buyers and sellers into platform-specific markets. 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 64.5% 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: 17% of attempts settle, agents carry 16% of online GMV and merchants retain 14% after platform take; 20% of buyers hold a wallet and 65% of sellers answer a rail. Top failure: no instrument at 30%.78