A2A Moves to the Linux Foundation: Who Sets the Standard for AI Agents?

Google has handed its Agent-to-Agent protocol to the Linux Foundation, effectively opening the standards debate for AI agents. 150 organizations, Azure and AWS on board — what the move means and what it implies for your business.

In June 2025, Google handed its Agent-to-Agent protocol (A2A) to the Linux Foundation. What sounds like an administrative formality is a strategic signal: A2A leaves Google's sphere of influence and becomes a vendor-neutral standard — governed like Linux, MySQL, or Kubernetes. One year later, the project counts over 150 supporting organizations according to the Linux Foundation and is firmly embedded in Azure AI Foundry and Amazon Bedrock AgentCore.

For companies building AI agents and automation today, a practical question arises: which protocols will prevail — and what should you build on when setting up agentic systems?

What Is A2A — and Why Did It Move to the Linux Foundation?

A2A addresses a problem that becomes more pressing as AI matures: how do AI agents from different vendors communicate with each other? Without a shared standard, every connection requires a custom integration — costly, maintenance-heavy, and hard to scale. A2A standardizes this contract: agents can discover each other, negotiate capabilities, and securely delegate tasks — regardless of whether they run on AWS, Azure, or Google Cloud. Google transferred the protocol to the Linux Foundation to ensure neutral governance and broader industry consensus.

MCP and A2A: Two Protocols, One Architecture

A2A is not the only protocol in the field. Anthropic's Model Context Protocol (MCP) solves a different problem — and complements A2A rather than competing with it: MCP connects an agent to tools and data sources (databases, APIs, internal systems), while A2A connects agents to each other. Together they form the infrastructure layer for complex automations — similar to how TCP/IP and HTTP in networking do not compete but serve different layers.

• MCP (Anthropic): Agent ↔ Tools — connects agents to tools, databases, and internal systems
• A2A (Linux Foundation): Agent ↔ Agent — standardizes communication and task delegation between different agents
• In production multi-agent architectures, both protocols are typically used simultaneously

Where A2A Runs in Production Today

A2A is not a lab protocol. Version 1.0 (April 2026) brings cryptographic Agent Cards for identity verification and five production-ready SDKs (Python, JavaScript, Java, Go, .NET). Founding members include AWS, Microsoft, SAP, Salesforce, and ServiceNow — for SAP customers in the mid-market, this is a relevant signal for the medium-term roadmap. In practice, A2A is already running in supply chain automations, financial services, and IT operations. The extension protocol AP2 (Agent Payments Protocol) also adds secure, auditable transactions — relevant for industries with compliance requirements.

What This Means for Your Business Today

Mid-sized companies do not need to implement A2A immediately. However, anyone who strategically evaluates AI agent projects should keep platform decisions in mind: in grown IT environments with multiple systems, agent interoperability quickly becomes a quality criterion. Those who bet on proprietary silos create tomorrow's integration backlog. What is needed is not an immediate switch, but an informed technology decision.

• Check: Does your AI platform provider support A2A — or is it part of the Linux Foundation initiative?
• Separate MCP and A2A in your planning: tools integration (MCP) vs. agent-to-agent coordination (A2A)
• SAP environments: SAP is a founding member — A2A integration in existing SAP landscapes is realistic in the medium term
• For new projects: Azure AI Foundry and AWS Bedrock AgentCore already include native A2A support
• Avoid proprietary point-to-point integrations between agents — the open standard is now available

An AI protocol under Linux Foundation governance means the same for companies as an open networking standard: no vendor lock-in,

no proprietary dead end — but also no obligation to rebuild everything immediately.

Those who want to understand how AI agents fit into the overall AI tier architecture — from simple assistants to fully automated workflows — will find a structured overview in the article AI Tiers Explained: From ChatGPT to AI Agents and Automated Workflows by NoviCogi Technologies.