Google’s AI agent protocol is becoming the language for digital labor

Google’s recently released Agent2Agent protocol has emerged as a top contender to facilitate communication and collaboration between AI agents in models and enterprises.

Microsoft this week adopted the A2A protocol in Copilot Studio and Azure AI Foundry for developers to build shared agents that can interact with each other.

Introduced last month, Google’ open-source A2A protocol allows networks of agents to structurally set goals, reason, take action, and return results across clouds, enterprises, and data silos, said Yina Arenas, vice president of product for Azure AI Foundry, and Bas Brekelmans, chief technology officer of Copilot Studio, in a joint blog entry.

“As customers scale these systems, interoperability is no longer optional,” the blog authors said.

[ Agentic AI: Ongoing coverage of its impact on the enterprise ]

Like Microsoft, AI services company Glean, which provides AI-based collaborative tools, is in the process of implementing A2A, Steve Calvert, software engineer at Glean, told Computerworld.

“Many enterprises are still trying to adopt agents, and A2A is really laying the foundation for the next generation of AI-powered collaboration,” Calvert said.

Calvert said the A2A open protocol helps agents communicate and work together, regardless of which vendor or system they’re built on. Today’s agents are designed to handle specific or a class of tasks, but A2A will change that, he said.

“For example, a sales agent supporting prospecting work, or an engineering agent generating postmortems — in the future, customers will want those agents to talk to one another, so they can solve even more complex problems,” Calvert said.

A2A is emerging at a time Anthropic’s MCP (model context protocol), another agent-to-agent protocol, is gaining steam.

A2A enables the orchestration of multiple agents, while MCP gives agents access to tools, Calvert said.

In practice, this means that agents will work with one another using A2A and interact with other systems using MCP, Calvert said.

A2A and MCP don’t compete, as they fill different requirements and “kind of need each other for the agents to work together,” said Bob Parker, senior vice president at IDC.

AI agents need middleware to understand and contextualize data in complex large-scale deployments, which is where MCP fits in. A2A is “more of asynchronous communication between agents themselves” that is emerging as a default protocol in the AI industry, Parker said.

For example, agents within M365 can talk directly within apps or services using A2A, which is a “more immediate mechanism to benefit productivity,” Parker said.

“I think long term, the MCP is going to be perhaps more impactful when you get outside of what you have on your desktop. When you start to talk about enterprise applications… MCP will be important,” Parker said.

Salesforce, which developed the A2A standard with Google, is also moving forward with the technology.

The A2A standard “allows AI agents to work together seamlessly across Agentforce and other ecosystems to turn disconnected capabilities into orchestrated [implementations],” said Gary Lerhaupt, vice president of product architecture at Salesforce, in an email statement to Computerworld.

The development of A2A is being watched closely by Read.AI, which has products that integrate with productivity and process tools from Salesforce, Atlassian, Google, and Slack. “There’s a lot of buzz around A2A — and MCP too, for that matter — and like so much of AI, these protocols are going to be transformative and the changes will happen fast,” Elliott Waldron, co-founder and vice president of data science, told Computerworld.

But it’s still early in terms of implementation for Read.AI, Waldron said. “We require a tighter integration of different sources to achieve the highest level of fidelity for our products. So we’re not moving yet, but we are definitely paying close attention,” Waldron said.