A2A Protocol v1.0入門 — 150組織採用のエージェント間通信標準をPythonで実装する Google's A2A inter-agent communication protocol, adopted by 150 organizations, reached its…

The Agent2Agent (A2A) protocol, first announced by Google in April 2025, reportedly reached its v1.0 milestone around its one-year anniversary in April 2026, and now counts roughly 150 organizations among its backers. For developers building systems where multiple AI agents must cooperate, A2A matters because it offers a vendor-neutral way for agents written in different frameworks and hosted by different companies to discover one another, exchange messages, and coordinate on work without bespoke point-to-point integration. The roster of supporters cited in announcements has included AWS, Microsoft, Salesforce, SAP, and ServiceNow, among many others.

A2A addresses a problem that has become more visible as agentic systems proliferate. Tools such as Anthropic's Model Context Protocol (MCP) standardize how a single agent connects to external data sources and tools, but they were not designed to govern communication between independent agents. A2A is positioned as a complement to MCP rather than a competitor: where MCP connects an agent to its tools and context, A2A connects agents to other agents. The two are frequently described as covering adjacent layers of the same stack, and many implementations use both together.

A central concept in A2A is the Agent Card, a JSON document typically served at a well-known location such as /.well-known/agent-card.json. The card advertises an agent's identity, the skills it offers, its supported input and output modalities, and its authentication requirements. A client agent fetches this card to decide whether and how to delegate a piece of work, which makes capability discovery a first-class feature rather than something hard-coded by integrators.

Communication is organized around tasks. A client sends a message to a remote agent, which creates a task that moves through a defined lifecycle, generally including states such as submitted, working, input-required, completed, failed, and canceled. Results are returned as artifacts, and intermediate updates can be streamed. For interactions that run for seconds, minutes, or longer, A2A supports streaming via Server-Sent Events and asynchronous push notifications, so a calling agent does not have to hold a connection open for the entire duration of a long task.

On the wire, A2A has historically used JSON-RPC 2.0 over HTTPS. The v1.0 specification is reported to formalize multiple transport options, including gRPC and HTTP-based JSON, with the aim of fitting more naturally into existing enterprise infrastructure. Security is handled through standard web mechanisms, with agents declaring their authentication schemes in the Agent Card and relying on transport-layer protections, which lets organizations apply familiar identity and access controls rather than inventing new ones.

For Python developers, the practical entry point is the official a2a-sdk package, which can be installed from PyPI. A minimal server implementation involves defining an agent's skills, exposing an Agent Card, and registering a handler that receives incoming messages and returns artifacts or status updates. The SDK provides the request routing, task management, and serialization needed to conform to the specification, so most of the developer's effort goes into the agent's actual logic. A client, in turn, resolves a remote agent's card, constructs a message, and either awaits a final result or subscribes to a stream of updates. Because the protocol is transport- and framework-neutral, an agent built with one toolkit can call an agent built with another, which is the interoperability benefit the standard is meant to deliver.

The governance context is worth noting. In mid-2025, Google handed A2A to the Linux Foundation, placing it under open, neutral stewardship alongside other infrastructure projects. That move appears intended to reassure adopters that no single vendor controls the specification, which is often a prerequisite for broad enterprise commitment. The growth from an initial set of launch partners to a much larger group of organizations is consistent with that strategy, though adoption figures cited in marketing materials should be read as claims of support rather than proof of production deployment.

For teams evaluating A2A, the sensible framing is that it standardizes one specific layer: how autonomous agents talk to each other. It does not replace orchestration frameworks, agent runtimes, or model providers, and it works best when paired with complementary standards like MCP. Readers planning to build on v1.0 should consult the official specification and SDK documentation directly, since details such as transport support and lifecycle semantics continue to evolve as the project matures.