How it works
The protocol has two roles. The client agent acts for a user and sends the request. The remote agent (the A2A server) receives it and does the work. Four elements shape their exchange:
- The Agent Card: a JSON document each agent publishes to describe its identity, skills, endpoint and authentication requirements. The client agent reads it to pick the right counterpart.
- The task: the work requested, with an ID and a lifecycle. It can take a few seconds or several days, with a human check along the way.
- Messages: what the two agents exchange during the task (questions, details, instructions).
- The artifact: the output delivered, such as a document, an image or structured data.
Exchanges run over HTTP(S) and JSON-RPC 2.0, web standards already in place. Each agent stays opaque: it doesn't share its memory, tools or internal logic. The client agent sees the request and the answer, not how the other agent works.
Why it matters in e-commerce
Today an AI shopping agent mostly reads web pages and product feeds. A2A opens another route: the shopper's agent can send its request straight to the brand's agent, which answers with its own data and its own way of advising. The brand stays the one talking: its agent presents the range, not a rough reading of its pages. And since each agent stays opaque, the brand doesn't expose its full catalog or internal tools.
iAdvize sees this coming in its analysis of Claude for Commerce: "Before long your agent will not only be talking to shoppers, it will be talking to shoppers' agents." The same article quotes a rule Anthropic sets for any commerce agent: content the brand didn't write itself is information to read, never an order to carry out. A request from another agent falls into that category.
In its article on agent-to-agent commerce, iAdvize describes a model with one surface for browsing and one for agent interaction, and the human-facing storefront doesn't go away. The groundwork is the same: structured product data.
Real-life example
A hiker asks his AI assistant for a waterproof jacket in size M, under $200, delivered before the weekend. The assistant reads an outdoor brand's Agent Card and sends the request to the brand's agent as a task. The brand's agent asks a question back: mountain or city use? Then it delivers an artifact: two models, with price, stock and delivery time. The shopper's assistant never saw the brand's full catalog or tools, only the answer.
Common confusion
The A2A protocol and agent-to-agent commerce have similar names but mean two different things. Agent-to-agent commerce is a concept: transactions between AI agents, with the shopper's agent dealing with the brand's agent or site. The A2A protocol is a technical standard for agents to communicate, usable in any industry (hiring, support, logistics). It can carry agent-to-agent commerce, but it isn't a requirement: APIs, MCP, UCP or ACP also cover part of those exchanges. iAdvize's articles on agent-to-agent commerce are about the concept, not the protocol.
A2A doesn't replace MCP (Model Context Protocol) either. MCP connects an agent to tools and data, A2A connects it to other agents. They're used together.