A2A protocol goes production
An agent‑to‑agent (A2A) protocol has crossed an adoption threshold — more than 150 organisations and placement in major cloud platforms — and is already being used in enterprise production. If A2A sticks, it could standardise how agents talk to each other the way service contracts did for microservices, making it easier to swap components and cut bespoke glue code. But standards alone won’t fix authority, provenance or liability disputes between agents — those semantic gaps remain open. (prnewswire.com).
Most artificial intelligence agents still talk to each other like companies that never agreed on email: every connection is a custom integration, and every new vendor adds another adapter. On April 9, 2026, the Linux Foundation said its Agent2Agent protocol had passed 150 supporting organizations and was already running in enterprise production. (prnewswire.com) Agent2Agent means exactly what it sounds like: one software agent can discover another, ask what it can do, hand it a task, and get work back in a standard format. The official documentation describes it as a common language for agents built by different vendors and frameworks. (a2a-protocol.org) Google introduced the protocol on April 9, 2025, with backing from more than 50 partners including Atlassian, Box, MongoDB, PayPal, Salesforce, SAP, ServiceNow, Workday, Accenture, Deloitte, PwC, TCS, and Wipro. The pitch was simple: if agents are going to book laptops, help customer service, or plan supply chains, they need a way to work across separate enterprise systems. (developers.googleblog.com) In June 2025, Google moved the project into the Linux Foundation, which matters because companies trust neutral standards bodies more than single-vendor promises. The Linux Foundation said the handoff was meant to keep Agent2Agent vendor-neutral and open to contributions across the industry. (linuxfoundation.org) The mechanics are closer to a service directory than to a chatbot conversation. Amazon Web Services says each agent can publish a capability manifest that lists what it can do, what inputs it needs, and operational details like version and latency, so other agents can match tasks to the right specialist. (docs.aws.amazon.com) That changes the architecture. Instead of one central orchestrator deciding every step, Amazon Web Services says Agent2Agent lets agents discover peers, negotiate tasks, and share context directly through a lightweight JavaScript Object Notation protocol, which is the plain-text data format used across the web. (docs.aws.amazon.com) The new part in this week’s announcement is not the idea but the placement. The Linux Foundation said Microsoft has integrated Agent2Agent into Azure AI Foundry and Copilot Studio, while Amazon Web Services added support through Amazon Bedrock AgentCore Runtime. (prnewswire.com) Microsoft’s own training material now teaches developers how to use Agent2Agent for agent discovery, direct communication, and coordinated task execution across remote agents. When a cloud platform starts shipping tutorials and exercises around a protocol, that usually means it is being treated as something developers are expected to build with, not just watch. (learn.microsoft.com) The project also says version 1.0 added multi-protocol support, multi-tenancy, updated security flows, and Signed Agent Cards for cryptographic identity checks. In plain English, that is the difference between a lab demo and a system a bank, insurer, or operations team might actually connect to production software. (prnewswire.com) Google has been careful to say Agent2Agent does not replace Model Context Protocol, which is Anthropic’s standard for connecting an agent to tools and data sources. The official documentation frames the split as tools on one side and other agents on the other, which is a useful distinction if both standards keep spreading. (developers.googleblog.com, a2a-protocol.org) What this does not solve is the hard human part. A standard message format can tell one agent how to ask another for a task, but it cannot settle whether the second agent had authority to act, whether its data was trustworthy, or who is liable when one agent makes a bad decision on another company’s behalf; those questions sit above the protocol layer, even if the plumbing is finally starting to standardize. (a2a-protocol.org, prnewswire.com)
