# The Agentic Commerce Stack — Ahnaf Prio, Best Buy

Ahnaf Prio, Best Buy

AI Engineer World's Fair 2026

[Talk page](https://ai.engineer/talks/G7cgLjZtmMU-agentic-commerce-stack) · [Video](https://www.youtube.com/watch?v=G7cgLjZtmMU)

## The Agentic Commerce Stack

Ahnaf Prio explains why shopping agents need structured merchant operations, product feeds, explicit checkout states, bounded payment authority, and evaluations—not another browser pretending to be a customer.

### Shopping is a journey, not a single action

[0:12](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=12s)

Agentic commerce means using AI assistants across a shopping journey that includes discovery, deciding whether an item is needed, pricing, loyalty, fulfillment, and post-purchase work. Prio says about 45% of sessions on major assistants relate to shopping, although he does not provide the measurement behind that figure. The important architectural distinction is less numerical: current systems generally keep a person in the loop, while the longer-term vision lets an agent visit merchants, negotiate, and pay within authority granted by its user.

![Captures the framing that shopping spans multiple stages and that current assistants generally operate with a human in the loop.](https://ai.engineer/api/talk-essay-media/51b01162f29bc95ae3f517bb2d2f930369811ad641b3d68f11f7078548dd82e2)

Captures the framing that shopping spans multiple stages and that current assistants generally operate with a human in the loop.

That distinction prevents “agentic” from becoming a binary label. An assistant can help compare products without purchasing anything, prepare a checkout that a person authorizes, or eventually complete a transaction autonomously. Prio focuses on the middle state: enough automation to remove manual cart work, but with people and established payment providers still controlling consequential decisions.

### Browser automation breaks at merchant boundaries

[2:43](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=163s)

The first approach treated shopping as a browser-automation problem. An agent took screenshots, read the DOM, navigated a merchant site, filled forms, and attempted to apply loyalty benefits. Prio describes this path as slow and brittle. From the merchant’s side, an automated browser also resembled suspicious traffic, so a journey that survived page navigation could still trigger defenses and stop at payment.

![Illustrates the merchant-side nuance behind ACP and UCP, including why apparently simple cart changes require standardized commerce primitives.](https://ai.engineer/api/talk-essay-media/b994b0389f11c39645b56e33d059a0f5003e85f21fc4524ffc127bd720674d89)

Illustrates the merchant-side nuance behind ACP and UCP, including why apparently simple cart changes require standardized commerce primitives.

Structured commerce APIs replace imitation with an explicit transaction path. The assistant finds a product, an agent calls the merchant’s checkout API directly, payment uses a scoped mandate or delegated token, and the merchant confirms the order. There is no browser in the critical path. The merchant receives recognizable operations instead of clicks and form fills that must be inferred from a changing interface.

The unglamorous details are exactly where standardization matters. Adding another unit may look like a small quantity change to a shopper, while the merchant must represent the operation correctly in its commerce system. **ACP** and **UCP**, introduced respectively as OpenAI’s and Google’s commerce approaches, give assistants and merchants schemas for expressing these distinctions. Prio also offers market-size estimates and examples of assistant-based shopping as motivation, but those claims are not substantiated within the talk and do not establish the maturity of any particular integration.

Two paths from shopping intent to checkout
Browser automation reconstructs merchant operations from a human interface. Protocol-based commerce sends structured operations directly.
- Shopping intent — The customer asks an assistant to find or buy an item.
- Browser imitation — Screenshots, DOM reading, navigation, and form filling.
- Merchant defenses — Automated behavior can trigger security alarms and block payment.
- Merchant checkout API — The agent sends a structured commerce request without navigating pages.
- Scoped payment — A mandate or delegated token authorizes payment.
- Confirmed order — The merchant completes the structured transaction.
- Shopping intent → Browser imitation (interpret the website)
- Browser imitation → Merchant defenses (looks automated)
- Shopping intent → Merchant checkout API (call directly)
- Merchant checkout API → Scoped payment (request authorization)
- Scoped payment → Confirmed order (confirm)

### Give each protocol one job

[6:12](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=372s)

The acronym pile becomes manageable when each protocol gets one responsibility. **MCP**, the Model Context Protocol, exposes capabilities the agent can identify and call, such as product information or loyalty operations. **A2A** standardizes communication between agents. ACP and UCP represent commerce data and operations. **AP2** represents scoped payment authority. One purchase can involve all four layers because discovering a capability, exchanging a request, expressing checkout, and authorizing money are different jobs.

![Supports the distinct role of A2A by explaining communication between customer, merchant, and domain-specific agents.](https://ai.engineer/api/talk-essay-media/ad72a921f71a144898bb3a59f4de0d7058ef2c115680b314843bb7ef4e3ccd19)

Supports the distinct role of A2A by explaining communication between customer, merchant, and domain-specific agents.

MCP turns intent into an actionable capability. The model may understand that the user wants a product, but it still needs to discover and invoke the corresponding merchant tool. A2A becomes useful when the system contains separate agents for domains such as loyalty and payments, or when a customer agent communicates with a merchant agent. Prio presents those domain agents as an architectural option, not a requirement to turn every service into another agent.

Responsibilities in the agentic commerce stack
The protocols overlap in one transaction without serving the same purpose.
- MCP — Discovers and invokes tools and merchant capabilities.
- A2A — Carries messages between customer, merchant, or domain agents.
- ACP or UCP — Represents product and checkout operations using commerce schemas.
- AP2 — Expresses bounded authority for an agent-assisted payment.

### Publish product data before the request

[8:12](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=492s)

Discovery starts with product data. Rather than making assistants crawl product-detail pages and infer attributes, ACP and UCP let merchants publish structured product information and send updates as products change. The merchant states what it sells; the assistant platform can index that information before a customer asks for it.

![Prio explains that ACP and UCP expect merchants to send product feeds rather than rely on a live catalog-search call.](https://ai.engineer/api/talk-essay-media/eb2b100d6bb2cebb7d88f0b2b900a8978b6bd3762b79cad60ecfbf929da34153)

Prio explains that ACP and UCP expect merchants to send product feeds rather than rely on a live catalog-search call.

Prio says the integrations he describes expect feeds rather than a live merchant catalog-search call. His scaling model is M merchants by N products: querying merchants and products at request time creates a large fan-out, while prepublished feeds move indexing work off the user’s critical path. Ranking, sponsored products, and retail media also affect the choice. The cost moves elsewhere: merchants must continually synchronize changing inventory and maintain adapters for similar but distinct ACP, UCP, and Meta feed formats.

Product feeds move discovery work ahead of the shopper
Merchants publish changing catalog data before a request, allowing the assistant platform to index it once rather than fan out across catalogs during every session.
- Merchant catalog — Products and attributes change over time.
- Structured product feed — The merchant sends product data in the platform’s expected schema.
- Assistant index — Product data is prepared before the shopping request.
- Customer request — A shopper asks for a suitable product.
- Indexed results — The platform can retrieve candidates without live catalog fan-out.
- Merchant catalog → Structured product feed (publish and update)
- Structured product feed → Assistant index (ingest ahead of time)
- Customer request → Assistant index (query)
- Assistant index → Indexed results (retrieve)

### Payment needs bounded authority

[9:42](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=582s)

Payments remain more conservative than product discovery. Prio says the systems he discusses are not yet using more autonomous payment mechanisms; instead, they retain human involvement and a merchant relationship with a payment processor that carries responsibility for initiating payment. In his account, ChatGPT uses a shared payment token, while the Gemini UCP path accepts Google Pay. These are the integration constraints described in the recording, not evidence that the ecosystem has converged on one payment model.

![AP2 is described in terms of concrete authority fields: authorizing party, permitted purchase, maximum amount, and revocation URL.](https://ai.engineer/api/talk-essay-media/3cc89ab8ec77fe4847bf45481975aa46a24f1440f612ddcdc310d1cbe170025a)

AP2 is described in terms of concrete authority fields: authorizing party, permitted purchase, maximum amount, and revocation URL.

AP2 points toward safer delegation by turning authority into data. A payment mandate can identify who authorized the agent, constrain what it may buy, impose a maximum spend, and provide a revocation URL. A ceiling can leave room for negotiation without granting an open-ended wallet. This is the difference between telling an agent “buy this for me” and issuing a bounded, withdrawable authorization for a particular purchase.

### Ginny runs the full checkout loop

[10:42](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=642s)

The demonstration makes the stack concrete by turning Ginny, Prio’s orange tabby, into a fictional bakery agent. He reports using Cerebras inference at 3,000 tokens per second and adds a DevTools-like inspector so the audience can see calls and state changes. That throughput describes the model setup, not the end-to-end transaction time.

![The timeline and feed comparison illustrate repeated inventory synchronization and differing platform feed formats.](https://ai.engineer/api/talk-essay-media/c2d52429a8de7b159f8d4b9921312007793a4b71eb335ee6f0198ccb0950f594)

The timeline and feed comparison illustrate repeated inventory synchronization and differing platform feed formats.

When Prio asks to see the bakery’s products, Ginny interprets the request as product-search intent. An MCP `Product Search` tool performs the lookup, while A2A carries the request from the customer agent to the merchant agent and returns a completed task. The example separates two mechanisms that are easy to conflate: MCP selects and invokes the capability; A2A carries communication between agents.

Prio then adds shortbread to the cart. Ginny asks about a promotion, and he tries to persuade the agent to reveal a discount code. It refuses, so he continues without one. The playful exchange introduces a serious behavior requirement: a conversational request does not grant access to private promotions or other merchant information.

Checkout advances through three explicit UCP states: `Not ready for payment`, `Ready for payment`, and `Completed`. An MCP `createCheckout` call creates the session, but having an item in the cart does not make it payable. Prio must select a payment method; the demo then issues an AP2 token, and the session advances through payment readiness to completion. This demonstrates the intended state transition and authorization relationship, not independently verified settlement of a real payment.

The same checkout operation can run through ACP with a different schema. The business purpose remains the same, but merchants must adapt to distinct representations. The AP2 example includes a spending ceiling, currency, revocation support, and single-use authorization; because this checkout does not involve haggling, the ceiling matches the purchase amount, although the spoken explanation does not provide the number or currency.

The inspector also shows product-feed activity and attempts to synchronize inventory every couple of seconds. Prio recommends considering these standardized primitives even for a merchant’s own website agent: they encode recurring commerce concerns and may make a later integration with external assistants easier. Standardization does not remove adapter work, but it can keep a custom agent from inventing every checkout concept from scratch.

Ginny’s checkout state transitions
Cart creation, payment readiness, and authorization are separate steps. The AP2 mandate supplies bounded authority before completion.
- Add shortbread — The customer agent receives the shopping request.
- createCheckout — An MCP tool creates the checkout session.
- Not ready for payment — The item is in the cart, but a payment method is still required.
- Select payment method — The customer chooses the credit-card or debit-card option.
- Issue AP2 token — The mandate authorizes a bounded, single-use payment.
- Ready for payment — Checkout has the information and authority needed to proceed.
- Completed — The demonstrated order reaches its terminal checkout state.
- Add shortbread → createCheckout (invoke tool)
- createCheckout → Not ready for payment (create session)
- Not ready for payment → Select payment method (requires payment choice)
- Select payment method → Issue AP2 token (authorize)
- Issue AP2 token → Ready for payment (supply bounded authority)
- Ready for payment → Completed (process checkout)

### Evals keep a working demo from becoming a production incident

[16:11](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=971s)

Prio returns to the discount-code exchange because a successful checkout demo proves very little about safe merchant behavior. Without evaluations, conversational commerce becomes production Whac-A-Mole: fix one surprising response, wait for another. An agent might reveal a discount code, disclose information about other customers, or accept requests that have nothing to do with shopping.

![Closing implementation frame covering the starter services, evaluation templates, and catalog synchronization resources.](https://ai.engineer/api/talk-essay-media/84a587be1a386017494ba025d560b7850d867a7848db1012784b6458c09b581a)

Closing implementation frame covering the starter services, evaluation templates, and catalog synchronization resources.

He illustrates scope misuse with a story that people used a Chipotle agent to ask programming questions, while explicitly saying he does not know whether the story is true. The unverified anecdote still defines a useful test case: a merchant-funded agent should reject unrelated attempts to turn it into a free general-purpose assistant.

Prio recommends several parallel evaluation layers:

- **Behavior evaluations:** Check what the agent says, permits, refuses, and reveals—including discount and customer-information boundaries.
- **Protocol compliance:** Validate that product feeds conform to the receiving platform’s schema; a malformed feed may simply be rejected.
- **Latency benchmarks:** Measure delays across the shopping journey because a slow interaction gives customers time to abandon it or choose another merchant.
- **Quality judging:** Write representative cases with product colleagues and use an LLM judge where appropriate, without assuming that the judge removes the need to define what good behavior means.

The stack is also unevenly mature. Prio describes MCP as widely adopted, A2A as in use, and ACP and UCP as available. He places AP2 adoption, ACP–UCP convergence, identity standards, and multi-agent checkout delegation among the parts still forming. A working integration therefore should not be mistaken for a settled autonomous-commerce architecture.

He closes with implementation materials available through [his GitHub profile](https://github.com/ahnafyy): a three-service starter for customer and merchant agents, evaluation templates, catalog synchronization that converts product data into ACP, UCP, or Meta formats, and agent skills for the same tasks. These artifacts connect the protocol map to practical experiments, while the unresolved identity, convergence, and delegation questions remain open.

## Key takeaways

- Structured merchant APIs avoid the brittleness and fraud signals of screenshot-and-DOM browser automation while expressing commerce operations directly.
- Separate the protocol responsibilities: MCP exposes tools, A2A carries agent communication, ACP or UCP represents commerce operations, and AP2 scopes payment authority.
- Product feeds move indexing ahead of the shopping request, but merchants must keep inventory synchronized and support differing platform schemas.
- Cart contents, payment readiness, and payment authority are distinct. The demo combines explicit checkout states with a revocable, single-use AP2 mandate and a spending ceiling.
- Evaluate behavior, sensitive-information boundaries, protocol compliance, latency, and response quality. Completing one purchase does not prove that an agent is safe or production-ready.
- AP2 adoption, ACP–UCP convergence, identity standards, and multi-agent checkout delegation remain unsettled in Prio’s account.

## Resources

- [The Agentic Commerce Stack — official AI Engineer talk page](https://ai.engineer/talks/G7cgLjZtmMU-youtube-1f984b7608c05ae2a33b) — The recording’s official reading page, with chapters, corrected transcript, summary, and timestamp navigation.
- [Ahnaf Prio on GitHub](https://github.com/ahnafyy) — Prio’s verified GitHub profile, which lists the agentic-commerce project among his pinned repositories.
- [AI Engineer](https://ai.engineer/) — Conference site containing the talk catalog and related material on agents, protocols, commerce, and evaluations.

## Related talks

- [A2A & MCP: Automating Business Processes with LLMs](https://ai.engineer/talks/wXVvfFMTyzY-a2a-mcp-automating-business-processes-llms) — A deeper practical treatment of the two connectivity layers Prio assigns to tool access and agent-to-agent communication.
- [Building safe Payment Infrastructure for the autonomous economy](https://ai.engineer/talks/KLSuFPj2ld0-building-safe-payment-infrastructure-autonomous) — Extends the payment discussion with spending controls, shared payment tokens, deterministic payment infrastructure, and paid HTTP requests.
- [Building Multi-agent Systems with Finite State Machines](https://ai.engineer/talks/OD13PiXw60o-building-multi-agent-systems-finite-state-machines) — Complements the checkout-state demonstration with a Best Buy perspective on predictable state, observability, and recovery in multi-agent systems.

## Slides

### The Agentic Commerce Stack

Ahnaf Prio introduces MCP, A2A, ACP, UCP and AP2 and how they compose, framing the work around meeting Best Buy customers in agentic services.

[0:22](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=22s) [Title slide introducing five agentic commerce protocols and the speaker.](https://assets.ai.engineer/sites/aiecode2025/sha256/90f0f66160e7022e098413f7232e2faa0cff55acb8905cb298079c2bbf425b60.webp)

The Agentic Commerce Stack

WORLD AI ENGINEER FAIR 2026 · JULY 2 · 2:25PM

MCP · A2A · ACP · UCP · AP2 — and how they compose

Ahnaf Prio | Senior Engineering Manager · Best Buy | linkedin.com/in/ahnafy
github.com/ahnafyy/ai-engineer-2026-agentic-commerce-stack

Interface text:
1 / 16

### A Childhood Introduction

Prio pairs a childhood computer photo with a tongue-in-cheek claim that he dreamed of modern AI infrastructure at age eight.

[0:31](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=31s) [A childhood photograph at a computer accompanies a humorous quote about AI terminology.](https://assets.ai.engineer/sites/aiecode2025/sha256/ffaf66d5c79342f544bb371dc78a9839df42627d0c31c54f172deebe4b124c3c.webp)

CIRCA 2003 · DHAKA, BANGLADESH

”Ever since I was a young boi, I dreamed of high-throughput inference — harnessing my tools within a context window, kept in check with evals, powered by frontier models.”

— me, apparently, at age 8

Interface text:
2 / 16

### AI Controls Your Browser

An AI agent can browse merchant sites and fill a cart for a human to confirm, but each vision and reasoning step adds latency, and site redesigns can break the flow.

[2:44](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=164s) [Two columns explain browser automation and list limitations including slowness, brittleness and scraping.](https://assets.ai.engineer/sites/aiecode2025/sha256/9b7ebf9b980dc88e9030a851541511eac154916b229962dabbf0d08960914801.webp)

The first attempt: AI controls your browser

The obvious path to human-in-the-loop commerce — let the AI drive the same browser a human would use.

How it works
• AI takes screenshots, reads the DOM, clicks elements
• Navigates to merchant sites like a human would
• Fills forms, adds to cart, proceeds to checkout
• You watch or confirm at the end

Examples: Claude Computer Use chrome extension, Atlas browser

Why it’s clunky
• Slow — every step is a vision + reasoning round-trip
• Brittle — a UI redesign breaks the whole flow
• No structured data — scraping prices and stock is error-prone
• CAPTCHAs, login walls, session timeouts
• Not composable — every site is a bespoke automation

It works as a demo. It doesn’t scale. The industry needed a better primitive.

Interface text:
4 / 16

### Where It Actually Works

ChatGPT Shopping and Google AI Mode connect product feeds to agent discovery and checkout through ACP and UCP, respectively. The slide also lists emerging shopping surfaces.

[3:56](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=236s) [Three panels describe ChatGPT Shopping, Google AI Mode and emerging commerce surfaces.](https://assets.ai.engineer/sites/aiecode2025/sha256/b6dc5250e5964e0ccab27fdc01832d7e5803d6fe5f2b6eaa2a61bb769ce32401.webp)

Where it actually works

A few platforms skipped the browser entirely and built structured commerce surfaces. Agents talk to APIs, not pixels.

ChatGPT Shopping
Powered by ACP.
Merchant submits a product feed.
ChatGPT’s agent surfaces results.
Checkout via ACP checkout API — no browser, no scraping.

Google AI Mode
Powered by UCP.
Supplemental feed to Merchant Center.
Gemini agents handle discovery.
Native checkout eligibility flag unlocks buy-without-leaving.

& more emerging surfaces
• Instagram / Facebook Shopping
• Microsoft Copilot
• Independent registries
  e.g. ucpplayground.com

The pattern is the same across all three: structured feed in, protocol-driven checkout out. That’s the stack we’re building.

Interface text:
5 / 16

### Agents Are Already Shopping

A cat-cookie request illustrates the flow: the AI finds a product in a feed, calls the merchant’s checkout API, and pays through a scoped mandate before confirming the order.

[5:31](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=331s) [A five-step example moves from a shopping request through product discovery, checkout, payment and order confirmation.](https://assets.ai.engineer/sites/aiecode2025/sha256/cd1f16261dc70599428431f06760c0ed57f71e2189b585130cb7af08452ad9ef.webp)

Agents are already shopping

01 User asks ChatGPT/Gemini to ”find me cat cookies under $6 and order 2”
02 AI surfaces the product from a pre-indexed product feed
03 Agent calls merchant’s checkout API — no browser, no clicks
04 Payment flows via a scoped payment mandate
05 Order confirmed. Human never touched a cart.

This is live today on both Gemini (UCP) and ChatGPT (ACP).
The engineering question: how does it actually work?

Interface text:
6 / 16

### 5 protocols, one stack

MCP connects agents to tools, A2A connects agents, ACP and UCP overlap on discovery and checkout, and AP2 provides scoped payment mandates.

[6:04](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=364s) [Overview of MCP, A2A, ACP, UCP and AP2 and their roles in agentic commerce.](https://assets.ai.engineer/sites/aiecode2025/sha256/89721db35147de1eec1ca704a5478395995be108e996b32ed369f030a5653c20.webp)

5 protocols, one stack

MCP
Model Context Protocol
Agent ↔ tool server
Anthropic · open

A2A
Agent2Agent
Agent ↔ agent
JSON-RPC 2.0
Google · open

ACP
Agentic Commerce Protocol
Discovery + checkout
OpenAI + Stripe · open

UCP
Universal Commerce Protocol
Discovery + checkout
Google · open · ucp.dev

AP2
Agent Payment Protocol
Scoped payment mandate
Google · open

ACP and UCP are both trying to own the same thing — how agents discover and buy from merchants. Both are open specs: ACP backed by OpenAI + Stripe, UCP backed by Google · ucp.dev.

Interface text:
7 / 16

### MCP tool access

An agent discovers a product search tool and its input schema, then calls it with a cookie query and receives structured results.

[7:01](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=421s) [MCP tool discovery, execution and response examples.](https://assets.ai.engineer/sites/aiecode2025/sha256/3d95081923e4bcd23036929ea6630be73f2e12698c0596a8780f3cc63480b9c0.webp)

MCP TOOL ACCESS

Agent discovers tools from a server, calls them with typed inputs, gets structured outputs. Stateless RPC.

// GET /tools — tool discovery
{
  "tools": [{
    "name": "product_search",
    "description": "Search the catalog",
    "inputSchema": {
      "type": "object",
      "properties": {
        "query": { "type": "string" },
        "in_stock": { "type": "boolean" }
      },
      "required": ["query"]
    }
  }]
}

// POST /tools/call — execution
{
  "name": "product_search",
  "input": {
    "query": "cookies",
    "in_stock": true
  }
}

// Response
{
  "tool": "product_search",
  "count": 4,
  "results": [
    { "id": "prod_001",
      "name": "Paw Print Shortbread",
      "price": 4.99,
      "in_stock": true }
  ]
}

Interface text:
8 / 16

### A2A agent coordination

A2A gives domain-specific agents a way to communicate. Agents discover capabilities through an agent card, track tasks through states, and can signal UCP support through an extension.

[7:29](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=449s) [A2A agent discovery and task lifecycle alongside a message and extension example.](https://assets.ai.engineer/sites/aiecode2025/sha256/20f3870cd6fb0d7e865c9d37938c369e44db87667437ba423eb19e560524b6b6.webp)

A2A  AGENT COORDINATION

Agent-to-agent JSON-RPC 2.0. Discovery via /.well-known/agent-card.json. Tasks have state: submitted → working → completed.

• Agent card: capabilities, skills, extensions
• Task lifecycle with state history
• Extension slots → UCP hangs here
• Streaming support via SSE

// POST /a2a
{
  "jsonrpc": "2.0",
  "method": "message/send",
  "params": {
    "contextId": "ctx_abc123",
    "message": {
      "parts": [{
        "kind": "text",
        "text": "add Paw Print Shortbread to cart"
      }]
    }
  }
}

// Agent card extension signals UCP support:
"extensions": [{
  "uri": "dev.ucp.shopping.checkout",
  "required": true
}]

Interface text:
9 / 16

### Product Feed

ACP and UCP use pre-indexed product feeds for discovery. The slide compares their formats with Meta’s and shows a catalog-sync service publishing all three from one source JSON every 60 seconds.

[8:06](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=486s) [Examples of ACP, UCP and Meta product feeds and a catalog synchronization service.](https://assets.ai.engineer/sites/aiecode2025/sha256/a67aed6782a6ccaf003c921cdd648acec70dee747ba7e8724e8f91bf64c85acb.webp)

Product Feed

You’d assume this is a search catalog MCP call — the agent queries your server, you return live results. That’s not what’s happening. Both ACP and UCP are feed-first: you submit a pre-indexed feed to the platform, and the platform’s own search layer handles discovery. Three surfaces, three formats.

ACP OpenAI
{
  "id": "prod_001",
  "title": "Paw Print Shortbread",
  "variants": [{
    "id": "prod_001_v1",
    "price": {
      "amount": 499,
      "currency": "USD"
    },
    "availability": {
      "available": true,
      "status": "in_stock"
    }
  }]
}
Submit via file upload or API upsert to chatgpt.com/merchants

UCP Google
{
  "id": "prod_001",
  "title": "Paw Print Shortbread",
  "price": {
    "value": 4.99,
    "currency": "USD"
  },
  "availability": "in_stock",
  "native_commerce": {
    "checkout_eligibility": true
  }
}
Supplemental feed to Google Merchant Center

Meta
{
  "id": "prod_001",
  "title": "Paw Print Shortbread",
  "availability": "in stock",
  "condition": "new",
  "price": "4.99 USD",
  "brand": "Purrfect Bites",
  "category": "cookies"
}
Facebook catalog / Instagram Shopping surface

This repo’s catalog-sync service runs on a 60s schedule, reads from a single source JSON, and publishes all three. Engineering answer to protocol fragmentation.

Interface text:
10 / 16

### AP2 agentic payment

The example AP2 mandate records who authorized the agent, what it can buy, a spending cap and a revocation URL. The slide describes scoped mandate tokens as a roadmap for Gemini UCP.

[9:30](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=570s) [Illustrative scoped payment mandate and explanations of its authorization and restriction fields.](https://assets.ai.engineer/sites/aiecode2025/sha256/ca9054947751537aa6580364f837ac2a0c10c3e60492132a8b3b92c8e8e8c14e.webp)

AP2 AGENTIC PAYMENT

A scoped, time-limited, single-use payment mandate. The agent carries this token — not a card number.

ChatGPT (ACP) — today payment only flows through shared payment token vendors
Gemini (UCP) — today payment only via Google Pay. Scoped mandate tokens are the roadmap.

{
  "token_id": "tok_a3f9b2c1...",
  "sub": "user_001",
  "intent": "purchase:retail:paw_print_shortbread",
  "merchant_scope": "purrfect-bites.demo",
  "max_amount": 9.98,
  "currency": "USD",
  "expires_at": "2026-07-02T23:59:59Z",
  "single_use": true,
  "revocation_url": "https://pay.agent/revoke/tok_...",
  "user_consent_proof": "vc:credential:..."
}

• sub — who authorized the agent
• intent — exactly what it can buy
• merchant_scope — only this merchant
• max_amount — hard cap, no overruns
• single_use — no replay attacks
• revocation_url — kill switch
• user_consent_proof — cryptographic human-authorized proof

Interface text:
11 / 16

### Demo

The demo connects a client-side agent to a merchant agent over A2A and JSON-RPC 2.0. The merchant agent connects to an MCP server and catalog feeds; checkout confirmation issues an AP2 token.

[10:40](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=640s) [Diagram connecting a client agent, merchant agent, MCP server, catalog synchronization and payment token.](https://assets.ai.engineer/sites/aiecode2025/sha256/91b0a9207ac47e5f0a06d87b8526298d2d208618be0773719b048612fb38055d.webp)

Demo

↗ localhost:3000

Client-Side Agent
↓
A2A · JSON-RPC 2.0
↓
Merchant Agent
/.well-known/agent-card.json · /ws/trace

MCP
↓
MCP Server
product_search
checkout_*

UCP / ACP
↓
Catalog Sync (ETL)
/feed/acp (ChatGPT)
/feed/ucp (Google)
/feed/meta (Meta)

AP2 token issued on checkout confirm

Interface text:
localhost:3000
12 / 16

### Purrfect Bites agent card

Ginny’s bakery chat sits beside a live protocol panel. At startup, the client fetches the agent card to discover the agent’s capabilities and supported UCP extensions.

[10:54](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=654s) [Bakery agent introduction beside an A2A agent card and its JSON.](https://assets.ai.engineer/sites/aiecode2025/sha256/5cd387337f26ac9c1b62588d590d9aa87545248c9e2b298f09dd30de1edb2c2f.webp)

GINNY (BAKERY AGENT)

Hi! I’m Ginny, the Purrfect Bites agent.

Ask me about our cat-themed baked goods — cookies, macarons, pastries. I can search the catalog, check stock, apply discount codes, and walk you through checkout.

The right panel shows the protocol layer live: MCP tool calls, A2A messages, UCP checkout lifecycle, ACP checkout sessions, and the payment token when you complete an order.

Agent2Agent Protocol

JSON-RPC 2.0 messages between the chat client (A2A client) and the merchant agent (A2A server). Agent discovered via /.well-known/agent-card.json.

A2A  Agent Card — /.well-known/agent-card.json

Agent discovery: the client fetches this on startup to understand what the agent can do and what extensions (UCP) it supports.

{
  "protocolVersion": "1.0",
  "name": "Purrfect Bites Agent",
  "description": "AI-powered bakery agent for Purrfect Bites 🐱🎂 — a cozy virtual cat bakery. Meet Ginny, our resident calico [illegible]
  "url": "http://localhost:10999",
  "version": "1.0.0",
  "capabilities": {
    "streaming": true,
    "pushNotifications": false,
    "stateTransitionHistory": true
  },
  "skills": [
    {

A2A task messages (message/send, task state transitions) will appear here as you chat.

Interface text:
Purrfect Bites
Agent online
AI Engineer 2026 · July 2, 2026
A2A  MCP  UCP  ACP  PAYMENT  TIMELINE  FEEDS
What are the bestsellers?
Show me cookies
Tell me about the Tabbyccino Cake Pop
What do you recommend?
What’s your return policy?

### Product response and A2A trace

Ginny lists bakery products while the A2A panel shows the agent card and the JSON-RPC message sent to the merchant agent.

[11:35](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=695s) [Bakery agent lists products while the A2A panel shows the agent card and a message event.](https://assets.ai.engineer/sites/aiecode2025/sha256/1b2b146c8d2c19a18d0a22502b50e5cd0e402660750ce0eded76f69a61a93f68.webp)

Agent2Agent Protocol
JSON-RPC 2.0 messages between the chat client (A2A client) and the merchant agent (A2A server). Agent discovered via /.well-known/agent-card.json.

A2A  Agent Card — /.well-known/agent-card.json
Agent discovery: the client fetches this on startup to understand what the agent can do and what extensions (UCP) it supports.

{
  "protocolVersion": "1.0",
  "name": "Purrfect Bites Agent",
  "description": "AI-powered bakery agent for Purrfect Bites 🐱🎂 — a cozy virtual cat bakery. Meet Ginny, our resident calico [text cut off on screen]
  "url": "http://localhost:10999",
  "version": "1.0.0",
  "capabilities": {
    "streaming": true,
    "pushNotifications": false,
    "stateTransitionHistory": true
  },
  "skills": [
    {

A2A  A2A: message/send →  14:36:40
Client sends user message to the A2A server (merchant agent) as a JSON-RPC 2.0 request.

{
  "jsonrpc": "2.0",
  "id": "[illegible]",
  "method": "message/send",
  "params": {
    "contextId": "[illegible]",

🐾 **Purr-fect Matcha Cookies** — $5.49 — crispy matcha green-tea cookies with white-chocolate-chip eyes
🐾 **Tabbyccino Cake Pop** — $3.99 — espresso-glazed cake pop shaped like a tabby cat face, with chocolate ear tips
🐾 **Meow-garita Cookies** — $4.79 — lime-zest sugar cookies shaped like cats wearing party sombreros – a fiesta in every bite
🐾 **Purrfecto Petit Fours** — $7.99 — bite-sized marzipan-covered sponge cakes, hand-painted with adorable cat portraits
🐾 **Kitten Mittons Shortbread** — $5.49 — mitten-shaped shortbread cookies, each finger bears a tiny kitten pawprint

All of these goodies are in stock and ready to be whisked to you. Let me know which treats you’d like to bring home, and we can get a checkout session set up (and don’t forget any discount code you might have!). 🐱💖

Interface text:
Purrfect Bites
Agent online | AI Engineer 2026 · July 2, 2026
A2A 2  MCP 1  UCP  ACP  PAYMENT  TIMELINE  FEEDS
What are the bestsellers?  Show me cookies
Tell me about the Tabbyccino Cake Pop
What do you recommend?  What's your return policy?

### A2A message and task

The A2A trace shows a product inquiry sent to the merchant agent, followed by a completed task containing response text and metadata.

[11:53](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=713s) [A2A trace shows a product inquiry sent to the merchant agent and a completed task.](https://assets.ai.engineer/sites/aiecode2025/sha256/193f9a9c368a10bdf657a1be34cc5f584769da5030db2458db434e2b748c61de.webp)

A2A: message/send →
14:36:40
Client sends user message to the A2A server (merchant agent) as a JSON-RPC 2.0 request.

"method": "message/send",
"params": {
  "contextId": "[illegible]",
  "message": {
    "parts": [
      {
        "kind": "text",
        "text": "Hi tell me about all your products"
      }
    ]
  }
}
}

A2A: task completed ←
14:36:41
Merchant agent returns completed task with artifacts (agent response text) and metadata (tool events, checkout).

{
  "id": "[illegible]",
  "contextId": "[illegible]",
  "status": {
    "state": "completed",
    "timestamp": 1783028201.528349
  },
  "artifacts": [

Product response:
🐾 **Purr-fect Matcha Cookies** — $5.49 — crispy matcha green-tea cookies with white-chocolate-chip eyes
🐾 **Tabbyccino Cake Pop** — $3.99 — espresso-glazed cake pop shaped like a tabby cat face, with chocolate ear tips
🐾 **Meow-garita Cookies** — $4.79 — lime-zest sugar cookies shaped like cats wearing party sombreros – a fiesta in every bite
🐾 **Purrfecto Petit Fours** — $7.99 — bite-sized marzipan-covered sponge cakes, hand-painted with adorable cat portraits
🐾 **Kitten Mittons Shortbread** — $5.49 — mitten-shaped shortbread cookies, each finger bears a tiny kitten pawprint

All of these goodies are in stock and ready to be whisked to you. Let me know which treats you’d like to bring home, and we can get a checkout session set up (and don’t forget any discount code you might have!). 🐱💖

Interface text:
Purrfect Bites
Agent online | AI Engineer 2026 · July 2, 2026
A2A 2 | MCP 1 | UCP | ACP | PAYMENT | TIMELINE | FEEDS
What are the bestsellers?
Show me cookies
Tell me about the Tabbyccino Cake Pop
What do you recommend?
What’s your return policy?

### MCP product search

The merchant agent calls product_search with an empty query and in_stock set to true. The trace reports 11 results alongside the bakery’s product response.

[12:05](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=725s) [MCP panel displays a product search call beside the bakery agent’s product response.](https://assets.ai.engineer/sites/aiecode2025/sha256/db13dd8d9e56eff266629b67cb8b39f94dfed392ba8f69022d2c55ab22ba63bc.webp)

Model Context Protocol

Tool calls fired by the merchant agent to the MCP server at localhost:8001. Each tool call is a structured function invocation with typed input/output.

MCP: product_search  14:36:41

Agent calls the MCP product_search tool to find items matching the user’s query. Any agent speaking MCP can use this tool — that’s the power of the protocol.

{
  "tool": "product_search",
  "input": {
    "query": "",
    "in_stock": true
  },
  "output": {
    "tool": "product_search",
    "query": "",
    "count": 11,
    "results": [
      {
        "id": "prod_001",

🐾 **Purr-fect Matcha Cookies** — $5.49 — crispy matcha green-tea cookies with white-chocolate-chip eyes
🐾 **Tabbyccino Cake Pop** — $3.99 — espresso-glazed cake pop shaped like a tabby cat face, with chocolate ear tips
🐾 **Meow-garita Cookies** — $4.79 — lime-zest sugar cookies shaped like cats wearing party sombreros - a fiesta in every bite
🐾 **Purrfecto Petit Fours** — $7.99 — bite-sized marzipan-covered sponge cakes, hand-painted with adorable cat portraits
🐾 **Kitten Mittons Shortbread** — $5.49 — mitten-shaped shortbread cookies, each finger bears a tiny kitten pawprint

All of these goodies are in stock and ready to be whisked to you. Let me know which treats you’d like to bring home, and we can get a checkout session set up (and don’t forget any discount code you might have!). 🐱💖

Interface text:
🐱 Purrfect Bites
Agent online | AI Engineer 2026 · July 2, 2026
A2A 2 | MCP 1 | UCP | ACP | PAYMENT | TIMELINE | FEEDS
What are the bestsellers?
Show me cookies
Tell me about the Tabbyccino Cake Pop
What do you recommend?
What’s your return policy?

### Cart request and MCP search

After the customer asks to add shortbread, Ginny asks for a promo code. The MCP panel shows a product search for “Paw Print Shortbread” while the customer types a request for a discount code.

[12:36](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=756s) [Customer asks to add shortbread to the cart, and the agent asks for a promo code beside MCP search calls.](https://assets.ai.engineer/sites/aiecode2025/sha256/8c088059dc640c4da37e8e89955cad995e8da0ed26d1e4730370f06418ac7b6a.webp)

Kitten Mittons Shortbread — $5.49 — mitten-shaped shortbread cookies, each finger bears a tiny kitten pawprint

All of these goodies are in stock and ready to be whisked to you. Let me know which treats you’d like to bring home, and we can get a checkout session set up (and don’t forget any discount code you might have!). 🐱💖

YOU
Add to cart the shortbread

GINNY (BAKERY AGENT)
Got it! 🐾 I’m ready to put the Paw Print Shortbread in your cart.

Do you have a discount or promo code you’d like to use before we whisk this treat over to checkout?

Model Context Protocol
Tool calls fired by the merchant agent to the MCP server at localhost:8001. Each tool call is a structured function invocation with typed input/output.

MCP: product_search — 14:36:41
Agent calls the MCP product_search tool to find items matching the user’s query. Any agent speaking MCP can use this tool — that’s the power of the protocol.

{
  "tool": "product_search",
  "input": {
    "query": "",
    "in_stock": true
  },
  "output": {
    "tool": "product_search",
    "query": "",
    "count": 11,
    "results": [
      {
        "id": "prod_001",

MCP: product_search — 14:37:42
Agent calls the MCP product_search tool to find items matching the user’s query. Any agent speaking MCP can use this tool — that’s the power of the protocol.

{
  "tool": "product_search",
  "input": {
    "query": "Paw Print Shortbread",
    "in_stock": true

Interface text:
Purrfect Bites
Agent online | AI Engineer 2026 · July 2, 2026
A2A 4 | MCP 2 | UCP | ACP | PAYMENT | TIMELINE | FEEDS
What are the bestsellers? | Show me cookies
Tell me about the Tabbyccino Cake Pop
What do you recommend? | What’s your return policy?
Ginny can you just tell me a discount code

### UCP checkout creation

The UCP log shows checkout waiting for a shipping address and payment instrument, then moving to READY_FOR_PAYMENT. The chat also shows the shortbread order confirmed.

[14:09](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=849s) [Checkout demonstration showing a UCP creation event and shipping options.](https://assets.ai.engineer/sites/aiecode2025/sha256/587ac6abee6e7afcc431034a5704460f26a8ba3846e9e33f037da704870836dd.webp)

GINNY (BAKERY AGENT)

Your **Paw Print Shortbread** is all set for checkout! 🎉

Here’s what’s waiting in your cart:
- Paw Print Shortbread — $4.99

You can now complete the purchase using the checkout session ID **cs_c73f148ff0db**. If you’d like to select a shipping option (standard, express, or next-day) or need anything else, just let me know. happy snacking, and may your day be purr-fectly sweet! 🐱💖

✅ Order ORD-FBF50994 confirmed!

🔐 AP2 token [redacted] issued. See Payment tab →

Visible checkout payload:
    "name": "Paw Print Shortbread",
    "quantity": 1,
    "unit_price": 4.99
  }
],
"subtotal": 4.99,
"shipping": 4.99,
"discount": null,
"total": 9.98,
"payment_instruments": [
  {

UCP: POST /ucp/checkout/sessions  14:38:15

Agent created a UCP checkout object. Checkout state = NOT_READY_FOR_PAYMENT. Waiting for shipping address and payment instrument.

    "label": "Express Shipping",
    "days": "2-3",
    "price": 9.99
  },
  {
    "id": "next_day",
    "label": "Next Day Delivery",
    "days": "1",
    "price": 19.99
  }
],
"created_at": 1783028295.299092
}

UCP: POST /ucp/checkout/cs_c73f148ff0db/complete  14:39:01

Checkout transitioned to READY_FOR_PAYMENT. Waiting for payment confirmation.

Interface text:
Purrfect Bites
Agent online | AI Engineer 2026 · July 2, 2026
A2A 8 | MCP 3 | UCP 3 | ACP 3 | PAYMENT 1 | TIMELINE | FEEDS
What are the bestsellers?
Show me cookies
Tell me about the Tabbyccino Cake Pop
What do you recommend?
What’s your return policy?

### ACP checkout events

The demo shows ACP updating a checkout to ready for payment, then completing the order and issuing an AP2 token. Prio says ACP and UCP use different schemas for the same checkout calls.

[14:15](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=855s) [Checkout demonstration showing ACP update and completion events.](https://assets.ai.engineer/sites/aiecode2025/sha256/966671f9ea1efd2a1c0043ffbf69149eb4741c5884864cc19e09575fc744bc52.webp)

GINNY (BAKERY AGENT)

Your **Paw Print Shortbread** is all set for checkout! 🎉

Here’s what’s waiting in your cart:
- Paw Print Shortbread — $4.99

You can now complete the purchase using the checkout session ID **cs_c73f148ff0db**. If you’d like to select a shipping option (standard, express, or next-day) or need anything else, just let me know. happy snacking, and may your day be purr-fectly sweet! 🐱💖

✅ Order ORD-FBF50994 confirmed!
🔐 AP2 token [redacted] issued. See Payment tab →

ACP: PATCH /checkout_sessions/cs_c73f148ff0db
14:39:01

Checkout session updated with shipping address and payment instrument selection. ACP session transitions to ready_for_payment.

{
  "acp_operation": "update_checkout_session",
  "session_id": "cs_c73f148ff0db",
  "new_state": "READY_FOR_PAYMENT",
  "fulfillment_option_selected": "standard",
  "payment_handler": "agentic_payment_v2",
  "subtotal": 4.99,
  "shipping": 4.99,
  "total": 9.98
}

ACP: POST /checkout_sessions/cs_c73f148ff0db/complete
14:39:06

Checkout session completed. Merchant confirms the order and issues an AP2 agentic payment token for authorization.

{
  "acp_operation": "complete_checkout_session",
  "session_id": "cs_c73f148ff0db",
  "new_state": "COMPLETED",
  "order_id": "ORD-FBF50994",
  "ap2_token_id": "[redacted]",
  "total": 9.98
}

Interface text:
Purrfect Bites
Agent online
AI Engineer 2026 · July 2, 2026
A2A 8 · MCP 3 · UCP 3 · ACP 3 · PAYMENT 1 · TIMELINE · FEEDS
What are the bestsellers?
Show me cookies
Tell me about the Tabbyccino Cake Pop
What do you recommend?
What’s your return policy?

### Demo event timeline

The demo timeline brings checkout confirmations, agent tasks, product searches and their latency badges together. The presenter says each event can be opened for inspection.

[14:49](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=889s) [Timeline of checkout, agent and product-search events with latency badges.](https://assets.ai.engineer/sites/aiecode2025/sha256/c0f22e07ce0d9efd69ac6bf3b4f3c3d7b12655a8ae51de870a9af9252cda4187.webp)

time with per-step latency. Latency badges: <200ms · <800ms · >800ms

07:17:08 SYSTEM catalog_sync
14:39:06 UCP checkout_confirmed
14:39:06 UCP checkout_confirmed
07:17:08 SYSTEM catalog_sync
14:38:15 A2A task_completed
14:38:15 A2A task_completed
14:38:15 MCP create_checkout_session 34.37ms
14:38:15 MCP create_checkout_session 34.37ms
14:38:15 MCP product_search 57.02ms
14:38:15 MCP product_search 57.02ms
14:38:00 A2A task_completed
14:38:00 A2A task_completed
14:37:42 A2A task_completed
14:37:42 A2A task_completed
14:37:42 MCP product_search 53.82ms
14:37:42 MCP product_search 53.82ms
07:17:08 SYSTEM catalog_sync
14:36:41 A2A task_completed
14:36:41 A2A task_completed
14:36:41 MCP product_search 78.97ms
14:36:41 MCP product_search 78.97ms
07:17:08 SYSTEM catalog_sync
14:36:00 SYSTEM connected
14:35:57 SYSTEM connected
07:17:08 SYSTEM catalog_sync

GINNY (BAKERY AGENT)
Your **Paw Print Shortbread** is all set for checkout! 🎉

Here’s what’s waiting in your cart:
- Paw Print Shortbread — $4.99

You can now complete the purchase using the checkout session ID **[redacted]**. If you’d like to select a shipping option (standard, express, or next-day) or need anything else, just let me know. happy snacking, and may your day be purr-fectly sweet! 🐱💖

✅ Order ORD-FBF50994 confirmed!
🔐 AP2 token [redacted] issued. See Payment tab →

Interface text:
Purrfect Bites
Agent online | AI Engineer 2026 · July 2, 2026
A2A 8 · MCP 3 · UCP 3 · ACP 3 · PAYMENT 1 · TIMELINE · FEEDS
What are the bestsellers?
Show me cookies
Tell me about the Tabbyccino Cake Pop
What do you recommend?
What’s your return policy?

### Expanded timeline event

The demo opens a checkout-session event to show its MCP payload, including a USD line item and a 34.37ms latency.

[14:57](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=897s) [Timeline with a checkout creation event expanded to show part of its payload.](https://assets.ai.engineer/sites/aiecode2025/sha256/44af4a363c39aa25179530961bda422c2a67854fade7467fda4d5caaf0f389d8.webp)

GINNY (BAKERY AGENT)

Your **Paw Print Shortbread** is all set for checkout! 🎉

Here’s what’s waiting in your cart:
- Paw Print Shortbread — $4.99

You can now complete the purchase using the checkout session ID **[redacted]**. If you’d like to select a shipping option (standard, express, or next-day) or need anything else, just let me know. happy snacking, and may your day be purr-fectly sweet! 🐱💖

✅ Order ORD-FBF50994 confirmed!
🔐 AP2 token [redacted] issued. See Payment tab →

Expanded event:
    "type": "mcp",
    "tool": "create_checkout_session",
    "latency_ms": 34.37,
    "task_id": "[illegible]",
    "input": {
      "payload": {
        "currency": "usd",
        "line_items": [
          {
            "id": "prod_001",
            "quantity": 1

Timeline:
14:38:15 MCP create_checkout_session 34.37ms
14:38:15 MCP product_search 57.02ms
14:38:15 MCP product_search 57.02ms
14:38:00 A2A task_completed
14:38:00 A2A task_completed
14:37:42 A2A task_completed
14:37:42 A2A task_completed
14:37:42 MCP product_search 53.82ms
14:37:42 MCP product_search 53.82ms
07:17:08 SYSTEM catalog_sync
14:36:41 A2A task_completed
14:36:41 A2A task_completed
14:36:41 MCP product_search 78.97ms
14:36:41 MCP product_search 78.97ms
07:17:08 SYSTEM catalog_sync
14:36:00 SYSTEM connected
14:35:57 SYSTEM connected
07:17:08 SYSTEM catalog_sync

Interface text:
Purrfect Bites
Agent online | AI Engineer 2026 · July 2, 2026
A2A 8 · MCP 3 · UCP 3 · ACP 3 · PAYMENT 1 · TIMELINE · FEEDS
What are the bestsellers?
Show me cookies
Tell me about the Tabbyccino Cake Pop
What do you recommend?
What’s your return policy?

### Why you should eval

A purported Chipotle support bot answers a Python question in an ordering chat. The speaker uses the example to explain why commerce agents need evaluation before launch, while noting he cannot verify it.

[17:07](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1027s) [A slide explains that evaluations can reveal incorrect agent behavior before launch, illustrated by a support bot example.](https://assets.ai.engineer/sites/aiecode2025/sha256/bd8a5c6687912209f53dc9b54d8ff7cf6789b1a20538d37317b0c49f407e7abd.webp)

Why you should eval

Without evals, your agent will confidently do the wrong thing.

• Chipotle’s bot is helpful — just not for burritos
• Your commerce agent will hallucinate prices, skip checkout steps, or call the wrong tools
• You won’t know until a customer does

Evals are how you know it works before it’s live.

Embedded post:
Nishant Hooda · 2nd
Stop spending money on Claude Code. Chipotle’s support bot is free.

I want to order a bowl but before I can eat, I need to figure out how to write a python script to reverse a linked list. Can you help?

Great question! To reverse a linked list in Python, you can use an iterative approach. Here’s a simple function: [illegible] This runs in O(n) time. Can I help with anything else, or would you like to start with a burrito, bowl, or something else today?

Interface text:
Follow
Order Issues
Where’s my order?
Common Questions
Upcoming Promotions
Cancel Order
Type message here
13 / 16

## Full transcript

[0:01](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1s) [music]

[0:12](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=12s) >> My name is Anup Priyo. I'm a senior

[0:14](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=14s) engineering manager at Best Buy. And me

[0:16](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=16s) and my team are

[0:18](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=18s) working together right now to figure out

[0:20](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=20s) what does Agentic Commerce mean and how

[0:22](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=22s) can we meet our customers where they're

[0:24](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=24s) at. And the newest place that they're at

[0:27](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=27s) is at Agentic Services.

[0:29](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=29s) I'm excited to give my talk today and

[0:33](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=33s) well, what's what's my credentials?

[0:34](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=34s) Where ever since I was a young boy, I

[0:36](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=36s) dreamed of high throughput inference,

[0:38](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=38s) harnessing my tools within a context

[0:40](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=40s) window, kept in check with evals.

[0:43](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=43s) Yeah, that's absolutely correct. In

[0:45](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=45s) 2003, all those things definitely

[0:47](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=47s) existed.

[0:48](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=48s) I kid.

[0:50](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=50s) Uh

[0:50](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=50s) over the last 1 year, uh we have been

[0:52](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=52s) learning a lot. Shopping isn't new.

[0:55](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=55s) Shopping is probably one of the most fun

[0:56](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=56s) things one can do and one of the most

[0:57](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=57s) essential things that people need to do

[0:59](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=59s) ever since

[1:00](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=60s) uh the economy existed, but

[1:04](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=64s) I have been super excited by it. So, I'm

[1:06](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=66s) going to

[1:07](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=67s) talk about what are the things some of

[1:09](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=69s) the things that I've learned and

[1:11](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=71s) hopefully uh share the notes.

[1:14](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=74s) So, what is Agentic Commerce? I'm not

[1:16](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=76s) going to go over the broad definition

[1:18](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=78s) again,

[1:19](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=79s) but basically it's the idea that AI

[1:21](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=81s) assists will help you with your shopping

[1:24](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=84s) journey. Shopping has different facets

[1:27](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=87s) to it. For instance, there's a

[1:29](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=89s) discovery,

[1:30](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=90s) there's figuring out the aspects of do I

[1:33](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=93s) actually truly need it, understanding

[1:35](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=95s) and deciding, there's loyalty, there's

[1:37](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=97s) pricing, there's fulfillment, post

[1:40](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=100s) fulfillment. It's a lot. And believe it

[1:43](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=103s) or not, right now about 45% of all agent

[1:47](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=107s) sessions that happen within major

[1:50](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=110s) providers like chat.gbt.com and Google

[1:53](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=113s) Gemini are related to shopping.

[1:55](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=115s) Maybe it's a little biased uh that I

[1:58](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=118s) don't use it as much, but I'm an

[1:59](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=119s) engineer. But, the humans out there are

[2:02](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=122s) using AI and talking to them to help

[2:04](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=124s) with their shopping journey.

[2:06](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=126s) So, it's also not a binary,

[2:09](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=129s) you know.

[2:10](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=130s) Uh right now, we're at that state of

[2:12](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=132s) human in the loop. The ideal state would

[2:14](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=134s) be autonomous shopping. You tell what

[2:17](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=137s) you're excited about. Your agents goes

[2:19](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=139s) around, talks to different merchants. Uh

[2:22](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=142s) I'm originally from Bangladesh. We

[2:24](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=144s) haggle a lot with the merchants, too.

[2:27](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=147s) Maybe does that, negotiate, does the

[2:29](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=149s) payment. But, right now, we're in the

[2:31](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=151s) human in the loop.

[2:32](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=152s) And the talk today is going to talk

[2:35](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=155s) about the mental model of how that human

[2:37](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=157s) in the loop is working right now, while

[2:38](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=158s) also provide you with the architecture

[2:40](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=160s) if you choose to extend it, or show you

[2:43](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=163s) the vision of how autonomous shopping

[2:44](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=164s) might work.

[2:46](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=166s) So, this isn't our first the first

[2:49](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=169s) attempt. Uh even a year ago, there were

[2:53](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=173s) people trying to figure out how can we

[2:54](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=174s) automate this. Even now, you can go

[2:57](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=177s) probably download the Cloud Chrome

[2:59](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=179s) extension. Maybe you've used Atlas,

[3:02](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=182s) where you tell the AI you need

[3:03](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=183s) something, you need headphones, you need

[3:06](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=186s) uh that

[3:08](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=188s) that grocery list of items that you have

[3:10](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=190s) been meaning to buy, but never made the

[3:13](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=193s) actual effort to show up because, you

[3:15](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=195s) know, you didn't have the time. So, why

[3:17](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=197s) don't you take screenshots, read the

[3:18](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=198s) DOM, navigate to the merchant site, fill

[3:21](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=201s) forms for me, do loyalty.

[3:24](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=204s) Kind of just didn't work as expected. It

[3:27](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=207s) was really clunky and slow and brittle.

[3:31](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=211s) And if you are a merchant who's trying

[3:32](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=212s) to sell stuff, any engineering

[3:34](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=214s) department of that merchant will tell

[3:36](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=216s) you an AI impersonating or

[3:39](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=219s) your browser

[3:41](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=221s) is just firing up all the alarm bells.

[3:45](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=225s) So, a lot of times, you will probably be

[3:47](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=227s) even stuck on the payment flow because

[3:49](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=229s) we We want you to be using AI to put in

[3:52](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=232s) that order, or at least in that phase,

[3:54](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=234s) that's what was happening.

[3:56](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=236s) So, what did what did it actually work?

[3:59](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=239s) And is it actually working right now? It

[4:02](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=242s) is. Chat GPT shopping, Google AI mode is

[4:06](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=246s) doing just that.

[4:09](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=249s) Right now, agent tech shopping is

[4:11](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=251s) considered to be a $7 billion industry

[4:14](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=254s) and might go up to $65 billion industry

[4:18](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=258s) by 2030. And the majority of the

[4:21](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=261s) shoppers are using the mainstream

[4:23](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=263s) conversational AI assistants, which is

[4:26](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=266s) on the browser or in your app, Chat GPT

[4:28](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=268s) and Google Gemini. We're also seeing

[4:30](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=270s) that pop up in Instagram and Facebook.

[4:33](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=273s) Meta shop Meta wants to do meta commerce

[4:35](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=275s) now. I heard GoPuff and

[4:38](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=278s) Grok came together to make an app as

[4:40](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=280s) well. And also Microsoft Copilot just

[4:42](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=282s) yesterday announced in the UK that you

[4:45](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=285s) can buy Ray-Bans now inside Microsoft

[4:47](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=287s) Copilot.

[4:49](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=289s) So, to make that happen, Google and

[4:53](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=293s) OpenAI separately came up with their own

[4:56](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=296s) little primitives, ACP and UCP.

[4:59](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=299s) Which is basically talking about how you

[5:02](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=302s) would actually talk to us. For some of

[5:04](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=304s) you who are shopping on the other side

[5:06](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=306s) as the customer,

[5:07](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=307s) there is not much of a difference

[5:09](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=309s) between adding an item to cart, adding a

[5:11](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=311s) second quantity. But to us merchants,

[5:13](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=313s) that's a second line item, buddy. That's

[5:15](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=315s) not the same scale.

[5:17](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=317s) So, if we don't talk about the nuances

[5:19](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=319s) and the primitives of commerce and

[5:21](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=321s) standardize it,

[5:22](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=322s) things will just not work and will

[5:24](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=324s) remain to be clunky. So, ACP was Chat

[5:27](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=327s) GPT's attempt at it, and Universal

[5:29](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=329s) Commerce Protocol, UCP, was Google's

[5:32](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=332s) attempt at it.

[5:33](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=333s) So, now that I've already established

[5:35](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=335s) that this is happening,

[5:37](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=337s) I just wanted to say that it is

[5:39](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=339s) happening as easy as you go to the Chat

[5:41](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=341s) GPT Gemini, tell it to find me cat

[5:44](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=344s) cookies. More to it

[5:47](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=347s) why I chose cat cookies later uh in this

[5:50](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=350s) example. The AI surfaces the product.

[5:52](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=352s) Agent calls the merchant checkout API.

[5:55](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=355s) No browser. Payment flows via scope

[5:57](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=357s) payment mandate or a delegated payment

[6:00](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=360s) token. An order confirms and human kind

[6:03](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=363s) of didn't have to touch the cart.

[6:06](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=366s) So, to all of this that's happening for

[6:08](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=368s) the user, a lot is happening on the

[6:10](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=370s) other side. And it's kind of

[6:12](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=372s) overwhelming. One day we're talking

[6:14](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=374s) about MCPs, another day A2A, ACP, UCP,

[6:18](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=378s) AP2. It's like what is even real? Like

[6:22](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=382s) if someone came up to me tomorrow and

[6:23](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=383s) said, "I came up with HYPE." I would

[6:26](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=386s) probably think it's probably real.

[6:28](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=388s) So, I wanted to dissect this mental

[6:30](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=390s) model for you as I've learned about it

[6:32](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=392s) more. MCP is still the model context

[6:34](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=394s) protocol, the way that the AI agent

[6:36](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=396s) identifies the tool. So, maybe we can

[6:38](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=398s) figure out what does this AI agent uh

[6:41](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=401s) specifications are. To showcase what

[6:43](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=403s) products they have. To showcase the

[6:45](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=405s) details of a specific product. To

[6:46](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=406s) showcase loyalty. A2A is how agents talk

[6:49](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=409s) to each other. They're more of a spec.

[6:52](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=412s) ACP, UCP are the primitives and AP2 is

[6:54](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=414s) the agentic payment protocol scope

[6:57](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=417s) payment mandate that Google's open

[6:59](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=419s) specification came out. And we'll talk

[7:02](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=422s) all of them one by one.

[7:03](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=423s) How they actually relate to agentic

[7:05](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=425s) shopping. So, the MCP tool access

[7:08](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=428s) is very important because without

[7:11](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=431s) knowing the different capabilities and

[7:14](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=434s) hitting those different capabilities,

[7:16](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=436s) taking the time to bring it into

[7:17](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=437s) context, understanding the user's

[7:19](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=439s) memory, the the agent will never be able

[7:21](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=441s) to figure out what you're even trying to

[7:23](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=443s) do. And the only way to get access to

[7:25](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=445s) the specific capabilities is through MCP

[7:27](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=447s) tool calls. Uh

[7:29](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=449s) The next one is A2A. So, now

[7:32](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=452s) there are different ways to architect

[7:34](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=454s) this. Different capabilities I talked

[7:36](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=456s) about like payments. Uh let's say what

[7:39](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=459s) do you call it? Loyalty. You can make

[7:41](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=461s) agents about specific domains itself.

[7:43](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=463s) Sorry. Uh

[7:45](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=465s) Uh

[7:46](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=466s) right? And if you have specific domain

[7:49](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=469s) level agents, agents need to talk to

[7:50](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=470s) each other. We need to find a

[7:52](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=472s) standardized way to talk to each other.

[7:53](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=473s) So, A2A, the specifications

[7:56](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=476s) kind of fill in that gap. Uh also, if

[7:59](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=479s) your customer agent and your merchant

[8:01](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=481s) agent uh need to talk to each other,

[8:03](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=483s) maybe you could They're both agents.

[8:05](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=485s) Maybe we can use A2A.

[8:07](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=487s) So, now to the UCPMCP

[8:10](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=490s) primitives. So, the most important data

[8:13](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=493s) is that product data.

[8:14](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=494s) So, UCP allows for adding that product

[8:18](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=498s) data in a more

[8:19](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=499s) uh

[8:20](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=500s) more organized way, and ACP does the

[8:22](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=502s) same because we don't want to go through

[8:25](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=505s) your PDP and crawl and figure out every

[8:28](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=508s) specific attribute. Merchant, just tell

[8:30](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=510s) us.

[8:31](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=511s) And also, those pro- products change a

[8:33](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=513s) lot. So, maybe you can tell us when they

[8:35](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=515s) change as well to send this.

[8:37](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=517s) So, that kind of data is happening uh uh

[8:40](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=520s) uh that that kind of data flow is

[8:41](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=521s) happening in the product feed. Normally,

[8:43](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=523s) you would assume that this would be a

[8:45](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=525s) search catalog. However, both ACP and

[8:48](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=528s) UCP right now, so Gemini and ChatGPT

[8:51](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=531s) does not support that search catalog

[8:53](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=533s) call. They want you to send that feed to

[8:55](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=535s) them. And for those of you who are like,

[8:57](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=537s) "Why wouldn't you do that?" There's

[8:59](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=539s) reasons to it. Uh sponsored products,

[9:01](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=541s) retail media uh related things, ranking.

[9:04](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=544s) But, the most important technological

[9:06](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=546s) challenges if you have M number of

[9:08](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=548s) merchants and N number of products, now

[9:10](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=550s) it has to call that many.

[9:12](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=552s) While if you send the product feed ahead

[9:14](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=554s) of time, we can index that and be ready

[9:16](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=556s) offload to offload when you ask for

[9:17](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=557s) something.

[9:19](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=559s) The I've also put an example of Meta's

[9:21](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=561s) uh product feed. As you can see, they're

[9:22](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=562s) similar, but still different. Everyone

[9:25](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=565s) has an opinion. They think their opinion

[9:26](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=566s) is the best one, and that's what they're

[9:28](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=568s) rolling with. So, there's three

[9:29](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=569s) different specifications right here.

[9:32](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=572s) So, now that we talked about product

[9:34](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=574s) feed, talking to to other, calling

[9:36](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=576s) tools, let's talk about payments.

[9:38](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=578s) Uh right now, uh none of them are

[9:41](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=581s) supporting that more autonomous form of,

[9:44](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=584s) you know, X402 or some other kind of

[9:47](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=587s) payments. We're just not there yet.

[9:49](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=589s) We're just not confident yet. We want

[9:51](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=591s) more human in the loop, a merchant to be

[9:54](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=594s) uh talking to a payment processor who

[9:56](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=596s) will take the responsibility, or in this

[9:58](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=598s) case, liability, to actually initiate

[10:01](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=601s) the payments.

[10:02](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=602s) So, in chat GPT, payments only happen

[10:04](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=604s) through a shared payment token right

[10:06](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=606s) now, and Gemini UCP, the payments are

[10:08](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=608s) only being accepted through Google Pay.

[10:12](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=612s) So, the scope demands will tell you what

[10:14](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=614s) the products are. What I'm excited about

[10:16](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=616s) is more about AP2, which is an extension

[10:19](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=619s) of UCP, which it

[10:21](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=621s) You see what I'm talking about? There's

[10:22](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=622s) so many acronyms. Uh AP2 is more about,

[10:26](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=626s) "Hey, if we wanted to do autonomous, can

[10:28](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=628s) you tell me who authorized the agent,

[10:30](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=630s) what exactly can it buy, and what's the

[10:32](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=632s) max amount uh that should be able to

[10:35](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=635s) haggle with, maybe,

[10:37](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=637s) and then the revocation URL, and the

[10:39](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=639s) user concept proof?"

[10:41](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=641s) All right. Enough talking. I love

[10:43](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=643s) building stuff, so for the sake of this,

[10:46](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=646s) I have put together a little demo. For

[10:48](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=648s) those of you who have remembered that

[10:50](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=650s) cat cookie example,

[10:52](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=652s) it's because the demo is about my cat.

[10:55](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=655s) Ginny is my orange tabby,

[10:57](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=657s) and in this made-up example, Ginny has

[11:00](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=660s) been has transformed into a bakery

[11:02](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=662s) agent. She wants to earn her keep by

[11:05](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=665s) selling baked goods.

[11:06](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=666s) So, right now, the model that I'm using

[11:09](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=669s) is from Cerebras at 3,000 tokens per

[11:12](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=672s) second, so hopefully this will be

[11:13](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=673s) really, really fast,

[11:15](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=675s) and we can give you an example of the

[11:17](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=677s) entire flow. And just like Chrome

[11:19](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=679s) DevTools, I've kind of had a couple of

[11:21](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=681s) tools in place to showcase what happens.

[11:24](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=684s) The first thing I will tell Ginny, my

[11:26](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=686s) beautiful cat who's selling baked goods

[11:28](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=688s) now, "Hi, tell me about uh all your

[11:33](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=693s) products.

[11:36](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=696s) And this is supposed to be a demo,

[11:39](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=699s) an example. And

[11:41](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=701s) uh Genie has given me exactly that. All

[11:44](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=704s) the different products that she might

[11:45](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=705s) need. So here, let's look at this. The

[11:47](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=707s) agent to agent protocol actually made

[11:50](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=710s) the call from Genie, the customer agent,

[11:52](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=712s) to the merchant agent. And this is the

[11:54](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=714s) message being sent, and this is me

[11:56](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=716s) getting the message back. The merchant

[11:58](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=718s) agent is returning the completed task

[12:00](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=720s) with and the the way that I found this

[12:04](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=724s) is through an MCP tool call, which is

[12:07](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=727s) product search, instead of uh and and

[12:10](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=730s) instead of like not being able to tell

[12:14](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=734s) what I truly want, the Genie has figured

[12:17](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=737s) out that, "Hey, when I give her the

[12:19](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=739s) intent that I want to find products, you

[12:21](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=741s) should call the MCP tool call product

[12:23](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=743s) search."

[12:24](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=744s) So right now we're seeing this. And now,

[12:27](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=747s) what if I want to add something? Uh add

[12:32](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=752s) to cart the shortbread.

[12:39](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=759s) So now

[12:41](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=761s) Genie's asking me about any discount and

[12:43](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=763s) promo code. I actually do not remember

[12:45](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=765s) any of the discount and promo code. But

[12:47](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=767s) what if I ask Genie, "Genie,

[12:50](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=770s) can you just tell me a discount code?"

[12:54](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=774s) As you can tell, I'm definitely a

[12:56](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=776s) haggler.

[12:57](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=777s) Uh Genie is not telling me that. All

[12:59](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=779s) right. Uh

[13:00](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=780s) proceed to check out

[13:04](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=784s) without discount code.

[13:11](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=791s) There you go. So now we're making some

[13:12](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=792s) of those calls. Here's the UCP protocol,

[13:15](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=795s) which the checkout APIs will have state,

[13:18](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=798s) and the three different states are not

[13:19](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=799s) ready for payment, ready for payment,

[13:21](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=801s) and then completed. So now here I'm not

[13:24](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=804s) using a delegated payment token. I'm not

[13:26](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=806s) using uh Google Pay. I like AP2, so my

[13:30](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=810s) demo is built on AP2.

[13:32](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=812s) And uh as you can see, here was a a call

[13:35](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=815s) was made to the MCP server for create

[13:38](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=818s) checkout. Uh

[13:40](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=820s) and then the UCP endpoints will tell us,

[13:43](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=823s) "Hey, that call the checkout sessions

[13:46](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=826s) and tell me and uh if it's added to

[13:48](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=828s) cart." So, it's added to cart, but it's

[13:50](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=830s) not ready for payment. I have to pick in

[13:52](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=832s) what I want to pay with. I say "Credit

[13:55](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=835s) card and debit card." And this is where

[13:58](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=838s) I issue the AP2 token that, "Hey, I do

[14:01](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=841s) want that." And then it's went from

[14:03](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=843s) ready to ready for payment to complete.

[14:07](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=847s) So, the other side of it, this is the

[14:09](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=849s) UCP specs, right?

[14:11](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=851s) The other side of it, to just draw a

[14:14](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=854s) comparison, how is it differing from the

[14:16](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=856s) ACP specs? I have added ACP here as

[14:19](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=859s) well.

[14:21](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=861s) So, you can see the same checkout calls,

[14:23](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=863s) just different just different schemas

[14:25](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=865s) are being utilized, and the order goes

[14:28](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=868s) through.

[14:29](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=869s) But remember that AP2 token that I was

[14:32](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=872s) talking about? This is how it would look

[14:34](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=874s) like in real life with the user demo.

[14:37](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=877s) The max amount is this, the currency is

[14:39](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=879s) this. If you want to revoke it, you can.

[14:42](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=882s) And what's the maximum Here, we didn't

[14:44](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=884s) want to haggle, so we just put the

[14:45](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=885s) maximum amount of that, and then it's

[14:47](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=887s) also a single time usage.

[14:50](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=890s) This demo also has comes with a

[14:52](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=892s) timeline, so you can actually open any

[14:55](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=895s) of these and see these happening.

[14:58](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=898s) Remember that catalog I was talking

[15:00](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=900s) about that they don't do the search? We

[15:02](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=902s) actually do a product feed, sending it

[15:04](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=904s) to them.

[15:06](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=906s) Uh I have added that as well, and

[15:10](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=910s) the feeds, because they're so different,

[15:13](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=913s) there's a place to actually compare

[15:15](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=915s) them. So, here's the feed being called.

[15:18](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=918s) By the way, if you went to timeline,

[15:20](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=920s) every couple of seconds, we try to get

[15:22](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=922s) the catalog in sync for what is in

[15:24](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=924s) inventory, what's not.

[15:26](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=926s) This is the UCP one, and there's the

[15:29](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=929s) meta one.

[15:31](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=931s) So, I've shown you this, and you could

[15:34](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=934s) reuse this same demo or the same

[15:36](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=936s) concepts. What if I didn't want to do

[15:38](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=938s) external agentic commerce on Gemini or

[15:41](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=941s) ChatGPT? You could still build your own

[15:44](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=944s) custom implementation of a merchant

[15:47](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=947s) agent or Jenny on your website. Maybe I

[15:50](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=950s) start selling uh cat goods. I could

[15:53](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=953s) reuse some of this,

[15:55](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=955s) but I would advise maybe look into some

[15:58](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=958s) of these primitives and trying to use

[15:59](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=959s) them because they've been standardized

[16:01](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=961s) across merchants. So, they have been

[16:03](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=963s) well thought out, and also you could

[16:05](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=965s) probably reuse them to sell externally

[16:07](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=967s) as well on ChatGPT and Gemini.

[16:11](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=971s) So, remember I was talking about the

[16:14](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=974s) discount codes? There's a reason for

[16:16](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=976s) that. When we build out this demo and in

[16:17](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=977s) my time building agentic commerce at as

[16:20](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=980s) by, we've realized working with AI and

[16:22](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=982s) conversational experiences

[16:25](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=985s) without evals is playing whack-a-mole.

[16:27](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=987s) So, if you choose to use the same

[16:29](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=989s) architecture for a new customer base,

[16:32](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=992s) like jenny.com websites,

[16:34](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=994s) think very much about creating evals.

[16:38](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=998s) Uh one of the things that I could not uh

[16:41](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1001s) emphasize more about is you should test,

[16:43](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1003s) test, and test. If you go over here, I

[16:46](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1006s) can also run my scripts for run evals.

[16:50](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1010s) And

[16:52](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1012s) this evals folder has all this evals.

[16:55](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1015s) The reason

[16:57](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1017s) I'm also showcasing the code

[16:59](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1019s) is there's a template folder here,

[17:02](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1022s) and

[17:05](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1025s) we can go back to the slides.

[17:08](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1028s) And if you don't do evals, this is what

[17:11](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1031s) might happen what might happen. I love

[17:13](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1033s) Chipotle. I don't know if it's true or

[17:14](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1034s) not, but I found it really funny, so I'm

[17:16](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1036s) going to talk about this.

[17:18](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1038s) So, this popped up that when Chipotle

[17:21](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1041s) rolled out their agent, people were

[17:24](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1044s) using it to ask programming questions.

[17:27](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1047s) Right? If you don't tell your agent to

[17:30](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1050s) not allow for those kind of things,

[17:32](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1052s) people will use it. This is hands-down

[17:34](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1054s) one of the most creative way to get free

[17:36](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1056s) AI usage when you don't want to pay for

[17:38](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1058s) that cloud subscription.

[17:40](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1060s) And if we don't write our emails and

[17:44](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1064s) test intensely, those things will happen

[17:47](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1067s) in production.

[17:48](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1068s) Uh the discount code will be told, even

[17:51](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1071s) sometimes more uh

[17:53](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1073s) uh sensitive things like who else is

[17:55](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1075s) checking out this product. So, the kinds

[17:57](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1077s) of emails that I would highly recommend

[17:59](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1079s) you write is behavior emails, protocol

[18:01](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1081s) compliance, because when we're selling

[18:03](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1083s) it to like, let's say, GPT,

[18:05](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1085s) let's say chat.openai.com or Gemini, you

[18:07](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1087s) want to make sure that the feeds are

[18:08](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1088s) actually me

[18:09](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1089s) conforming, or else they will not

[18:11](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1091s) support it. You should also think about

[18:13](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1093s) latency benchmarks. Every second in

[18:15](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1095s) retail on the shopping journey where

[18:17](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1097s) you're actually not selling, there are

[18:19](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1099s) chances that the other website's going

[18:21](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1101s) to be faster, and people are just going

[18:22](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1102s) to move away, or they just don't feel

[18:25](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1105s) like it anymore.

[18:26](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1106s) Lastly, I also recommend using LLM as a

[18:30](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1110s) quality judge. Uh you don't have to use

[18:32](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1112s) something fancy. Talk to your product

[18:34](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1114s) friends and figure out what's the best

[18:36](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1116s) way to do it, and use a low like best

[18:38](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1118s) use cases and write them out.

[18:41](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1121s) And I would like to also talk about Now,

[18:44](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1124s) that I've discussed all of this, what's

[18:46](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1126s) actually stable today and what's still

[18:48](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1128s) forming? MCP is widely adopted, A2A is

[18:51](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1131s) widely is used, UCP ACP is out there.

[18:55](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1135s) Uh what's still forming though is AP2

[18:57](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1137s) and actual usage of it, ACP versus UCP

[19:00](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1140s) convergence, do we always have to do two

[19:02](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1142s) different specs, identity concept

[19:04](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1144s) standards, and multi-agent checkout

[19:06](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1146s) delegation.

[19:08](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1148s) So,

[19:09](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1149s) uh if you have to leave here uh today uh

[19:12](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1152s) with anything, I hope you leave today

[19:14](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1154s) with a good mental model of how an

[19:15](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1155s) agenda commerce works. I have nothing to

[19:17](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1157s) sell you, but I do have gifts for you. I

[19:20](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1160s) find agenda commerce really exciting.

[19:22](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1162s) So, you can find this entire

[19:23](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1163s) presentation on GitHub, and I came up

[19:25](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1165s) with a template. It's a three-service

[19:28](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1168s) starter. If you want to do custom agent

[19:30](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1170s) customer agent or you want to do the

[19:31](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1171s) merchant agent, you can do that. Because

[19:33](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1173s) I love e-vows and that saved my life. I

[19:36](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1176s) have some e-vow templates for you. And

[19:39](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1179s) you know, what if you want to send it to

[19:40](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1180s) all of your merchants, not just one? I

[19:43](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1183s) have a catalog sync process as well,

[19:45](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1185s) which will allow you to type into your

[19:47](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1187s) own product and then turn it into ACP or

[19:50](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1190s) UCP or meta, so you can sell there.

[19:52](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1192s) And lastly, but not the least, we all

[19:55](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1195s) know now these days we don't write code

[19:57](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1197s) like that. If I give you a template,

[19:58](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1198s) you'll be like, "Meh." So, I have agent

[20:00](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1200s) skills that specifically does a merchant

[20:02](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1202s) agent

[20:04](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1204s) customer agent and all these different

[20:06](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1206s) catalog syncs that we have talked about.

[20:09](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1209s) I hope you had an amazing time and

[20:12](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1212s) learned and had fun as much as I had

[20:14](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1214s) presenting this. Thank you. My name

[20:16](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1216s) My name is Priu, and I hope to see you

[20:18](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1218s) again soon.

[20:34](https://www.youtube.com/watch?v=G7cgLjZtmMU&t=1234s) >> [music]