The technology and assumptions behind the purchasing chapters.
Our current scenario connects technology advances to purchasing experiences. Select a future to see what supports it and what remains unproven.
This is the scenario behind the current chapters. Early mechanisms have documented sources; later branches are conditional possibilities, not confirmed outcomes.
Find me an e-bike for my 8 km commute. I need local repairs and theft cover. Keep everything under $2,200. The assistant explains tradeoffs, then links you to the seller’s variants.
See what shopping could look like.
A conversation becomes the shop window → Watch the shopping experienceCan an agent compare current offers and catch exclusions across unfamiliar stores?
What would count: Repeat shopping tasks across stores, record successful outcomes and missed constraints.
Anthropic · Commerce Agents ↗Still unproven: A working architecture does not establish independent completion rates, merchant coverage or error recovery.
Research this advancement ↗Recent developments and ongoing signals shaping our scenarios. Sources inform the mechanisms; simulation percentages remain explicit assumptions.
A blueprint includes shopping and merchant agent reference implementations. Harnesses, tools and guardrails provide an implementation starting point.
Buyer and merchant agents could exchange complete offers. Our negotiation scenario also needs reliable execution and payment; the blueprint alone does not establish that future.
Explore: Sellers compete for a need →Independent cross-provider completion, negotiation quality and real stock commitments.
Measure successful orders, missed constraints, intervention and recovery on unseen stores.
Pages expose structured tools to a WebMCP-aware agent in the browser.
Structured store actions could let an agent select variants and complete carts across stores. The model treats WebMCP and browser operation as alternative routes, both requiring task reliability.
Explore: Your agent shops across stores →Browser support, merchant coverage, tool correctness and recovery from changed stock.
Visa describes cardholder authentication and passkeys for payment instructions. Agent-specific tokens and controls against authenticated payment instructions are documented.
A linked identity and payment token could turn a recommendation into a completed purchase. Our checkout mechanism needs both; reusable budgets and broad acceptance remain assumptions.
Explore: An agent buys →One-step linking across independent accounts and agents remains a hypothesis; identity is not purchase authority.
Acceptance, declines, refunds and enforceable limits across transactions; a reusable budget is a separate assumption.
Product results can appear in chat with details and merchant links.
Discovery and comparison move into a conversation. This supports the research chapter; completing a purchase depends on additional capabilities.
Explore: An agent researches →Accurate fit, current prices and coverage across the buyer’s options.
Shopping research can use previous conversations and memory to personalize recommendations.
Remembered preferences could reduce repeated briefs. The model combines memory with chat for a remembered shopper, and with physical signals, payments and recovery for an anticipated need.
Explore: The world notices the need →Separate explicit preferences from guesses; handle outdated needs, corrections, forgetting and privacy. Memory does not authorize spending.
ChatGPT search links support referral tracking. ChatGPT offers advertising placements. Organic product results are separate from ads.
The starting point of a shopping journey changes. Our model explores store handoffs and sponsored discovery separately; neither implies that an agent can buy.
Explore: An agent researches →A referral does not mean an agent operated the site or completed a purchase.
How sponsored exposure changes trust, shortlist quality and buyer choice.
Manually reviewed. New evidence is reviewed before changing a scenario or its assumptions.
Watch technologies, run a counterfactual or draft another chapter. Experiments here do not change the current findings above.
Inspect the same run as ideas or a dependency graph. Open a calculation to see why a result changed.
Choose assumptions above and run the simulation.
Technology coverage × task reliability gives usable capability coverage. Required capabilities multiply; alternative routes combine as 1 − product(1 − capability). This first model assumes independence. Shared platforms, failures and adoption can violate that assumption.
Trajectories interpolate linearly. Missing inputs produce bounds, never invented point estimates. Futures overlap; their percentages are not additive and are not probabilities that those futures will happen.
Ideas and research questions use a disclosed mechanism library. They are deterministic compositions of the selected shopping example and dependency results, not autonomous research or an LLM brainstorm. The graph comes from those same dependencies. The older numerical horizon model remains a separate diagnostic below.
Create a chapter brief from a result. Keep its assumptions, buyer flow, failure case and provenance together then play a matching interactive recipe.
Drafts are stored in this browser and can be downloaded. Saving does not publish a chapter. Seven mechanisms can generate playable previews using authored shopping surfaces; other ideas remain briefs.
Competition moves from owning the account to serving the mandate. The films are authored illustrations; these indicators are calculated.
| Delegated demand | Governance capacity | Adversarial loss |
|---|---|---|
| 19.8% | 60.0% | 8.2% |
capabilityLearning = 0.16 · inferenceCostDeflation = 0.22 · sellerAgentDiffusion = 0.15 · identityPortability = 0.8 · liabilityClarity = 0.52 · adversarialPressure = 0.2 · physicalExecutionDiffusion = 0.09 · collectiveMandatePortability = 0.28 · adaptiveSupplyDiffusion = 0.09
The storefront survives as a business boundary even when its pages disappear. The films are authored illustrations; these indicators are calculated.
| Delegated demand | Governance capacity | Adversarial loss |
|---|---|---|
| 37.0% | 60.0% | 6.9% |
capabilityLearning = 0.16 · inferenceCostDeflation = 0.22 · sellerAgentDiffusion = 0.11 · identityPortability = 0.8 · liabilityClarity = 0.52 · adversarialPressure = 0.2 · physicalExecutionDiffusion = 0.09 · collectiveMandatePortability = 0.28 · adaptiveSupplyDiffusion = 0.05
Autonomy scales through institutions that price and absorb exception risk. The films are authored illustrations; these indicators are calculated.
| Delegated demand | Governance capacity | Adversarial loss |
|---|---|---|
| 52.2% | 67.2% | 4.2% |
capabilityLearning = 0.16 · inferenceCostDeflation = 0.22 · sellerAgentDiffusion = 0.11 · identityPortability = 0.8 · liabilityClarity = 0.82 · adversarialPressure = 0.13 · physicalExecutionDiffusion = 0.09 · collectiveMandatePortability = 0.28 · adaptiveSupplyDiffusion = 0.05
Checkout becomes policy maintenance and the person moves to supervision. The films are authored illustrations; these indicators are calculated.
| Delegated demand | Governance capacity | Adversarial loss |
|---|---|---|
| 64.2% | 69.2% | 4.0% |
capabilityLearning = 0.17 · inferenceCostDeflation = 0.26 · sellerAgentDiffusion = 0.11 · identityPortability = 0.8 · liabilityClarity = 0.82 · adversarialPressure = 0.13 · physicalExecutionDiffusion = 0.09 · collectiveMandatePortability = 0.28 · adaptiveSupplyDiffusion = 0.05