Microsoft backs open A2A protocol to connect rival AI agents

AI agents are everywhere right now, but most of them are trapped inside their own software islands. An agent built on one cloud usually has no clue how to talk to an assistant running on another.

Microsoft wants to change that. On May 7, the company announced it is adopting the open Agent2Agent, or A2A, protocol across Azure AI Foundry and Copilot Studio.

In plain English, A2A gives AI agents a common language to share goals, pass state, invoke actions, and return results across different clouds, frameworks, and company borders.

Tearing down the walls between agents

Microsoft is rolling this out on top of some massive enterprise momentum. More than 70,000 organizations, including Epic, Fujitsu, and LG Electronics, currently build on Azure AI Foundry.

Its newer Agent Service picked up over 10,000 organizations in just four months. Meanwhile, Microsoft Copilot Studio is already tapped by over 230,000 organizations, including 90 percent of the Fortune 500.

All those companies want their agents to handle real multi-step workflows. But nobody wants to be locked into a single tech stack to get it done.

With A2A support, developers can build in frameworks they already know, such as LangChain or Semantic Kernel, and have those agents communicate cleanly.

Enterprise security is wired in by default. Calls travel through Microsoft Entra authentication, mutual TLS encryption, Azure AI Content Safety filters, and complete audit trails.

Copilot Studio agents will also gain the ability to call out to external agents hosted outside Microsoft. Microsoft has joined the GitHub A2A working group to help shape the spec, and a public preview inside Foundry and Copilot Studio is coming soon.

Developers can already check out a new Python sample in Semantic Kernel showing two local agents planning an itinerary and handling currency conversions without custom orchestration code.

Why it matters

While Anthropic's Model Context Protocol focuses on connecting an agent to data sources and tools, A2A is tackling how agents collaborate with each other.

For builders and enterprise teams, this is a huge relief. You can build specialized agents for individual departments or tasks without worrying about rewriting integration code every time you switch infrastructure.

We think Microsoft is playing this very smartly. By positioning Copilot as the UI for AI while pushing open protocols like A2A and MCP underneath, Microsoft keeps the user front and center without looking like a walled garden.

It forces competitors to either support open agent-to-agent standards or risk isolating their own tools.

That said, multi-agent systems operating across cloud boundaries carry real operational risks. Cascading failures, permission escalation, and latency bottlenecks get tricky fast once autonomous agents start delegating work to external machines. We will see how ready the enterprise world really is once the preview goes live.