A2Aプロトコル徹底解説：エージェント間通信の標準規格がもたらす「AIエージェントのインターネット」The A2A protocol defines a standardized communication spec for AI agents to…

The Agent2Agent (A2A) protocol is an open specification that lets autonomous AI agents discover one another, exchange messages, and coordinate on multi-step tasks regardless of the framework, vendor, or cloud that produced them. As enterprises deploy a growing number of narrowly specialized agents, the inability of those systems to talk to each other has emerged as a practical bottleneck. A2A addresses that gap by defining a common transport and message format, and its proponents frame it as a foundation for an open "internet of agents" in which heterogeneous systems can collaborate without bespoke integration work.

At a technical level, A2A is built on familiar web standards rather than a novel wire format. Communication typically runs over HTTP using JSON-RPC 2.0 for structured requests and responses, with Server-Sent Events used to stream incremental updates for long-running work. This choice is deliberate: by leaning on established protocols, A2A is intended to slot into existing infrastructure, load balancers, and authentication systems that engineering teams already operate, lowering the barrier to adoption.

The central discovery mechanism is the Agent Card, a machine-readable JSON document that an agent publishes to advertise itself. The card describes the agent's identity, the skills it offers, its endpoint URL, supported input and output formats, and the authentication schemes it requires. A card is commonly hosted at a well-known path so that a "client" agent can locate and inspect a "remote" agent before deciding to delegate work to it. This pattern resembles how services expose capability metadata elsewhere on the web, and it allows agents to negotiate what they can do for one another at runtime.

Interactions in A2A are organized around the concept of a task. A client agent sends a request, the remote agent creates a task with a unique identifier, and that task moves through defined states as it progresses. Because agent work can take seconds or hours, the protocol supports streaming partial results, polling for status, and push notifications so a client can be informed when a long task completes. The outputs of a task are returned as artifacts, and messages can carry multiple content types, or "parts," including text, structured data, and files. This design accommodates workflows in which one agent asks another to, for example, research a topic, draft a document, or process a dataset, then hand back the result.

A key point of context is how A2A relates to the Model Context Protocol (MCP), which is frequently discussed alongside it. The two are generally described as complementary rather than competing. MCP, introduced by Anthropic, standardizes how a single agent or model connects to external tools, data sources, and context, effectively acting as a bridge between a model and its resources. A2A instead standardizes how one agent communicates with another agent as a peer. In a typical architecture, an agent might use MCP to reach the tools it needs and A2A to collaborate with other agents, so many teams treat the two protocols as parts of the same interoperability stack.

A2A originated at Google, which announced it in 2025 with support from a broad set of technology and consulting companies, and it was subsequently contributed to the Linux Foundation to be governed as a vendor-neutral, open project. That governance move appears intended to reassure adopters that no single company controls the standard, which is often a prerequisite for wide industry uptake. The protocol has also emphasized enterprise concerns from the outset, including authentication, authorization, and the ability to run across organizational boundaries, reflecting the reality that agents from different companies may need to interact securely.

Several caveats are worth keeping in mind. Interoperability standards succeed only when enough implementers adopt them, and the practical value of A2A will depend on how consistently agent builders publish compatible Agent Cards and honor the task lifecycle. Questions around security, identity, trust between agents, and liability when autonomous systems act on each other's behalf remain active areas of work and are likely to evolve as real deployments accumulate. The ecosystem of software development kits, reference implementations, and testing tools is still maturing.

For developers evaluating the space, the pragmatic takeaway is that A2A and MCP together sketch a layered approach to agent systems, one connecting agents to tools and the other connecting agents to agents. Whether the "internet of agents" vision materializes at scale will depend on adoption, tooling, and the resolution of trust and security concerns, but the standardization effort itself lowers the cost of experimentation today.