Google Cloud Next 2025 - Rise of The Agent-2-Agent Protocol

Google Cloud Next 2025: The Rise of Agent-to-Agent Protocol

After attending day two of Google Cloud Next 2025, I've got some interesting insights to share about what looks to be the main theme of this year's conference: agents. While last year's buzzword counter might have been tracking "AI," this year it's all about "agents." The term was liberally sprinkled throughout the developer keynote and dominated conversations across the conference floor.

Agents Everywhere, But Are They Really?

I spent a significant portion of today exploring the expo area with fellow tech enthusiast, the YouTuber Prompt Engineering (if you're not familiar with his channel, you should definitely check it out). We recorded a podcast together and walked around evaluating the various agent demonstrations on display.

Our shared takeaway? We were both rather underwhelmed.

Most of what vendors were showcasing as "agents" were what we would more accurately describe as simple workflows or scripts—things like currency converters and other basic applications. While useful, these implementations fall short of the fully agentic solutions that have been promised. They're essentially clever LLM workflows rather than the autonomous agents that have been hyped as potential job replacements. Needless to say, AGI remains firmly on the distant horizon for all providers we encountered today.

 Google's Agent-to-Agent Protocol: What Is It?

The big announcement that caught my attention was Google's new agent-to-agent protocol (A2A), which was highlighted in yesterday's initial keynote and referenced repeatedly throughout today's developer sessions.

At its core, the A2A protocol aims to facilitate communication between different agents. Unlike existing frameworks such as MCP (which focuses on allowing an agent to access various tools and data), A2A operates at a different level—it's designed to enable sophisticated collaboration between agents.

Rather than simply using another agent as a tool or performing a basic handoff, A2A allows agents to find other relevant agents, negotiate tasks, specify output formats, determine languages, and share other important details. The protocol leverages standard technologies like HTTP, SSE, and JSON RPC to facilitate these interactions.

Google is emphasizing security by default with A2A—a point highlighted by a joke I heard at the conference: "Did you know the S in MCP stands for security?" Of course, there is no S in MCP, which was exactly the point being made about potential security concerns with that framework. Importantly, Google isn't positioning A2A as a replacement for MCP but as a complementary protocol operating at a different level of interaction.

Agent Discovery and the Future Agent Marketplace

One of the most intriguing aspects of A2A is its focus on agent discovery. This aligns with what has been a hot topic in agent framework companies for the past 18 months—the concept of "agent stores." The vision is that eventually, we'll have app store-like experiences where your agent can locate and select other specialized agents to accomplish specific tasks, potentially paying small fees for these services.

Google's implementation includes "agent cards" that describe what each agent can do, how it operates, and other relevant details. The protocol outlines various discovery methods, including public registries and private APIs, allowing agents to work with both public and internal private agents.

Technical Implementation and Partners

Google has made substantial resources available to showcase A2A, including a blog post, website, GitHub repository, and sample agents. The GitHub repo provides a conceptual overview covering agent cards, A2A servers, client interactions (which bear some resemblance to MCP), and key terminology around tasks, messages, and artifacts (outputs that agents pass back to users or other agents).

The company has announced 50 technology partners for the protocol, including model builder Cohere, agent framework company LangChain, and various consulting firms. However, there's a notable absence: Anthropic isn't on the list.

 Potential Concerns

This omission raises questions about the protocol's future. If Anthropic is planning its own "MCP version 2" that handles both agent-to-agent and tool interactions, we could end up with competing protocols—a scenario that wouldn't benefit the ecosystem.

While Google has a strong track record with protocols like gRPC, often making them open and eventually transferring them to independent governing bodies, the lack of certain key players is concerning. Beyond Anthropic, other notable agent frameworks like LlamaIndex and Paidantic AI are also missing from the partner list.

 What's Next?

The technology world is littered with abandoned or poorly adopted protocols. For A2A to succeed, broader industry buy-in seems essential. Google has indicated they're open to feedback on the protocol, so there's still opportunity for its evolution.

What do you think about the agent-to-agent protocol concept? Are you excited about the potential for agent marketplaces? What features would you like to see added to the protocol? Let me know in the comments—Google team members are likely to be reading!

Bloggers Note - This post is taken from a transcript of a video posted by  Sam Witteveen