Overview
Agentic commerce is becoming a practical retail channel rather than a speculative concept: conversational assistants can already discover products, initiate merchant checkout APIs, and complete purchases without manipulating a browser cart. The speaker presents a layered mental model for this ecosystem. MCP exposes merchant capabilities as callable tools; A2A enables specialized agents to communicate; ACP and UCP define commerce concepts such as product feeds and checkout sessions; and payment mechanisms—including shared tokens, Google Pay, and the emerging AP2 authorization model—control how transactions are approved. This architecture replaces screenshot-and-DOM automation, which proved slow, brittle, and likely to trigger merchant security systems. A working bakery-agent demo illustrates product discovery, cart creation, payment selection, authorization, and checkout completion across different schemas. However, the ecosystem remains fragmented: merchants may need multiple product-feed and checkout formats, major assistants currently prefer pre-ingested feeds over live catalog search, and autonomous payment delegation is immature. The operational lesson is equally important: conversational commerce must be governed by behavioral, protocol-compliance, latency, and quality evaluations. Without those controls, agents may reveal discounts or sensitive information, answer unrelated questions, violate partner schemas, or lose customers through slow responses.
Sections
Core Terms
The principal protocols and concepts used in the speaker's agentic-commerce model.
- Agentic commerce: the use of AI assistants to support multiple stages of a customer's shopping journey, potentially progressing from human-supervised assistance to autonomous purchasing.
- MCP, or Model Context Protocol: the tool-access layer through which an AI agent identifies and invokes merchant capabilities.
- A2A: a specification for communication between agents, including specialized domain agents and customer-to-merchant agent exchanges.
- ACP and UCP: commerce protocols that organize primitives and schemas for product data, checkout, and related merchant interactions; ACP is associated with ChatGPT and UCP with Google.
- AP2: an emerging agentic-payment extension associated with UCP that expresses scoped authorization, purchase constraints, spending limits, consent proof, and revocation.
- Product feed: merchant-supplied structured catalog data that an assistant provider can ingest and index before a shopper makes a request.
Strategic Implications
Higher-level conclusions derived from the architecture, demo, and operational recommendations.
- The decisive shift is not merely from websites to chat interfaces; it is from interfaces interpreted at runtime to explicit machine-readable commerce contracts.
- Product-feed ingestion gives assistant providers control over indexing, ranking, and retrieval, which may make agentic-commerce distribution resemble existing marketplace and retail-media ecosystems.
- Payment authorization is the practical boundary between assisted and autonomous commerce: discovery can be flexible, but purchasing requires precise authority, limits, evidence, and revocation.
- Because ACP, UCP, and Meta feeds differ, merchants benefit from maintaining one internal commerce model and translating it into external schemas rather than embedding provider conventions throughout their systems.
- In conversational retail, evaluation quality becomes part of commercial operations because agent behavior directly affects conversion, confidentiality, protocol acceptance, and customer trust.
Implementation Details
Specific architectural and operational details demonstrated or recommended in the talk.
- The demonstration uses a customer agent, a merchant agent, MCP tools, A2A messaging, checkout endpoints, product-feed synchronization, and an AP2-style authorization token.
- A product-search intent is mapped to an MCP product-search tool, while creating a purchase invokes a separate create-checkout tool.
- The UCP checkout session exposes the states 'not ready for payment,' 'ready for payment,' and 'completed.'
- The example AP2 authorization includes a maximum amount, currency, revocation mechanism, and single-use constraint.
- The demo periodically synchronizes catalog information to reflect inventory changes and can generate ACP, UCP, or Meta-compatible feeds from merchant product data.
- The recommended test suite covers behavioral boundaries, protocol conformance, latency, and LLM-judged response quality.
Key Comparisons
Contrasts among implementation approaches and protocol responsibilities.
- Browser automation interprets screenshots and the DOM and may trigger security controls; API-based agentic commerce uses structured product, checkout, and payment interfaces and is presented as more reliable.
- ACP and UCP support similar commerce flows, including checkout, but use different schemas and are associated with different assistant ecosystems.
- Live search would require runtime calls across many merchants and products; pre-ingested feeds allow providers to index catalog data in advance.
- Human-in-the-loop commerce keeps users involved in key decisions and payment authorization; autonomous shopping delegates negotiation and purchase execution within predefined constraints.