How industry efforts aim to help AI agents communicate better

If the tech industry is to be believed, an army of chatty computerized coworkers is headed to your office, in the form of AI agents. Within individual companies or industries, they may already be there.

But just like on-boarding a human hire usually involves many new logins, these digital drudges aren’t much use without access to at least some of the constellation of platforms that make up a modern workplace. And while they won’t be rehashing last night’s game at the watercooler, they do still need to talk to each other.

These are the kinds of problems that emerging industry standardization efforts are trying to tackle. Model Context Protocol (MCP), first introduced by Anthropic, is designed to plug agents into various software tools and data sources. Google’s Agent2Agent (A2A) protocol, which debuted this April, aims to let agents communicate, even across different vendors.

Like any attempt to coordinate among varying companies, these cross-compatibility efforts can involve a tricky balance of different needs and interests. And while MCP and A2A are among the most popular at the moment, other options exist, like IBM’s Agent Communication Protocol (ACP) or Oxford University’s Agora.

While it’s still very early days for these frameworks, experts said some form of communication and coordination will be needed for companies to realize their ambitions for agentic workplaces.

It's been three years since ChatGPT first thrust LLMs into the mainstream. As part of Morning Brew’s Quarter Century Project, we opted to take a look at why many businesses see these agent networks as the next big phase of AI in the workplace.

Rao Surapaneni, VP of Google Cloud’s business application platform, said A2A is meant to give agents a common language to communicate and share information.

“As a customer, when I’m deploying these platforms, how do I not reinvent the wheel and still be able to leverage expertise from different vendors and different agents?” Surapaneni explained to Tech Brew. “We designed the Agent2Agent protocol to leverage that, where the customer can retain their secure data, keep their business logic, but still expose the outcomes via agents.”

Beyond APIs: Right now, APIs are the “de facto mechanism for communication” between different digital tools, according to Matt McLarty, CTO of integration platform Boomi and co-author of a book about APIs. But McLarty said MCP offers a more consistent language geared specifically toward GenAI models.

“What MCP does is it adds just a little more structure around deriving the semantics of the services—having an understanding of the meaning of things—which you find in APIs, but it’s not mandated there,” McLarty said. “So it’s a way of providing a consistent way of presenting these remote services and data sources that is useful to LLMs, very much geared toward LLMs.”

In the year since MCP first launched, it’s gained momentum. A number of major foundation model vendors now support it, and public hubs list thousands of MCP servers for different tools and data sources. “MCP definitely seems to have the [head start] in terms of adoption for the tool usage,” João Moura, CEO of multiagent orchestration framework CrewAI, told us.

Agent speak: A2A is more nascent, but Google has enlisted about 150 partners to support the open-source framework, which aims to facilitate communication and information-sharing between agents.

Moura said mostly only “cutting-edge” clients are using A2A right now; for instance, one Fortune 500 client is using it to orchestrate ServiceNow agents. But it remains a burgeoning technology.

“A2A was kind of like a solution before a problem,” Moura said. “A lot of the companies out there, they just cracked prompting, and now they’re getting into agents…I don’t think there’s necessarily a huge mass of people trying to get agents to talk over the wire just yet.”

Surapaneni gave an example of onboarding a new hire, where a Workday agent might have to communicate with a ServiceNow agent to get the person a laptop, for instance. He said Google has also seen traction among companies using A2A to coordinate various supply-chain tools.

“In a supply-chain scenario, you have multiple vendors producing these agents. A vendor has a demand forecasting agent. Another vendor gives you the inventory management for their inventory, and then another vendor provides shipping times,” Surapaneni said. “So I’m seeing orchestration across these multiple agents in a mostly autonomous fashion.”

What the future holds: Based on the history of tech standardization efforts, it can be hard to predict which protocols will ultimately catch on, Moura said. McLarty said the way the protocols are currently delineated—agent-to-agent or agent-to-tool—might not make sense in the long run.

“If I were to place a bet right now, I think we’ll see MCP become fairly universal for cross-communication at runtime between all the agentic components,” McLarty said. “I think A2A will have its place…handling more of the trust negotiation.”

Surapaneni said Google is already starting to think about agent communication for consumers, imagining a world where everyone has a personal agent that navigates shopping and even engages with government services.

“This is the envisioned future, and some of these are possible now, and I expect these to become more and more real as we go through it,” he said.