No runtime fully closes the door, but no coordination layer makes every buyer instrument and seller rail work together either. Merchants repeat integrations, agents maintain adapters and cross-rail conversion remains partial. By 2029Q2 agents carry 19% of online GMV and 21% of attempts settle — better than the gardens, but three points of GMV and eight points of wallet adoption behind coordinated open rails.
No platform owns the whole purchase, yet no shared coordination layer makes every mandate and rail work together. Agents remain useful by maintaining adapters, reconciliation and fallbacks.
One constant: the selected household replenishment need and its human boundary. One variable: who can interpret that authority and where recourse lives.
By 2029 delegated commerce works with seams showing. The long-horizon question becomes whether seller agents can negotiate across those seams without turning exception handling into permanent infrastructure.
illustrative emerging mechanism · fragmented open web · 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 runtimes continue passing third-party payment challenges, but card credentials, wallets and seller rails only bridge in some combinations. Success reaches 9.5%, below the coordinated case, while UCP becomes the default seeded by storefront platforms. Buying agents start shipping their own adapter lists.7811
Agent vendors maintain compatibility matrices for rails, processors and storefront plugins. Sellers can stay on the open web, but supporting two protocols still feels like two projects. Half expose at least one rail; agents learn which combinations actually settle.37
Thirteen percent of attempts settle and agents carry 3% of online GMV. UCP leads settled volume, card tokens follow, and x402 and MPP each remain meaningful without becoming universal. Merchants increasingly ask whether a new integration adds reach or merely duplicates buyers they can already serve.711
More than half of sellers answer a rail, yet cross-instrument failures remain common enough that support teams can name them. Larger agents run their own routers; smaller agents depend on whichever adapters their runtime bundles. The web is open in policy and uneven in practice.28
A majority of sellers expose some machine-readable payment path. The remaining merchants see the demand but delay because integration, monitoring and reconciliation are still fragmented. Wallet adoption reaches 20%, five points behind the coordinated path at this point.29
Agent GMV approaches 10%. The best-resourced buying agents advertise how many storefront and payment combinations they can complete, turning routing coverage into a visible product feature. Merchants retain more than in a garden, while paying a quiet tax in duplicated operations.711
Success approaches 20%, but the buyer's missing instrument remains the largest single failure. Industry groups publish better profiles and conformance tests; adoption improves, though not fast enough to erase the gap created by years of partial bridges and duplicated seller work.38
Agents carry 19% of online GMV and 21% of attempts settle. No proprietary checkout owns the market, but no open rail feels universal either: UCP carries 23% of settled volume, card tokens 21%, platform checkout 20%, and x402 and MPP split most of the rest. The future remains contestable, not frictionless.678
Continue this institution into the 2040 story →
Open protocols remain available, but partial interoperability and duplicated integrations make adoption slower and less reliable. 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 64.8% 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: 22% of attempts settle, agents carry 20% of online GMV and merchants retain 19% after platform take; 27% of buyers hold a wallet and 65% of sellers answer a rail. Top failure: no instrument at 27%.78