Google's NEW Agent2Agent Protocol

Google’s new “agent-to-agent” (A2A) protocol is aimed at turning today’s tool-using AI agents into a network of collaborating agents that can discover each other, negotiate tasks, and exchange structured outputs—using web-friendly standards like HTTP, SSE, and JSON-RPC. The practical punchline is a shift from “one agent calls tools” to “agents coordinate with other agents,” with built-in emphasis on security and interoperability. Google positions A2A as complementary to MCP (Model Context Protocol), not a replacement, placing the protocol at a higher layer of agent-to-agent communication rather than tool access.

At Cloud Next 2025’s developer keynote, the word “agents” dominated the conversation, but demos in the expo area often looked more like polished LLM workflows—currency converters, small scripts, and other narrow utilities—than the fully agentic systems people associate with job-displacing automation. That context matters: A2A is presented as a more ambitious infrastructure step toward the “agent marketplace” future, where agents can locate other specialized agents, delegate work, and potentially get paid per task.

The core mechanism centers on “agent cards,” which function like profiles describing what an agent can do, how it works, and what format it returns results in. A2A envisions an agent client that can query a registry (or other discovery endpoints) to find relevant agents for a given task, then select which agent(s) to use. Communication is structured around tasks, messages, and “artifacts”—artifacts being the outputs passed back to users or forwarded between agents. In a typical flow, an input triggers discovery, then routing, then collaboration, all under a shared protocol.

Google also leans hard on security-by-default messaging, contrasting its approach with earlier assumptions around MCP’s security posture. The company’s materials—blog post, website, GitHub repo, sample agents, and a “agent developer kit”—spell out how A2A should work, including examples that suggest cross-language composition (e.g., an initial agent in JavaScript calling agents implemented in Python).

The strongest upside is clear: standardized discovery and collaboration could reduce fragmentation and make agent ecosystems more plug-and-play, enabling the app-store-like model for agents that many in the industry have been discussing for roughly a year and a half. The biggest concern is adoption risk. Google lists around 50 technology partners, including Coh and Langchain, but Anthropic is notably absent. If major model providers don’t align, the ecosystem could end up with competing protocols—an outcome that has historically plagued other standards. Google’s track record with open protocols like gRPC offers some reassurance, but the missing Anthropic name is the flashpoint.

In short, A2A is a bid to standardize agent collaboration and discovery at scale. Whether it becomes the backbone of an agent marketplace—or another abandoned protocol—may hinge on whether key players like Anthropic choose to participate and whether the industry can converge on one shared approach rather than splinter into incompatible layers.