AI Agent Interoperability Advances as A2A and MCP Gain Ground

Rivals are sharing the plumbing

AI agents are starting to work with one another, although the progress is slow. The shift isn't coming from one breakthrough. It's coming from a set of shared protocols that competing companies now support. Two of them do most of the work. Anthropic's Model Context Protocol (MCP) connects an agent to tools and data. Google's Agent2Agent protocol (A2A) lets one agent find another and hand it work. Both now sit under Linux Foundation governance, and every major cloud has adopted them. That makes this a question for enterprise buyers, not only for developers.

The main effect is on cost. Without a common standard, every pairing of an agent with a tool, or of one agent with another, needs its own custom code. One analysis points out that this maintenance burden grows quadratically as deployments expand.22 Shared interfaces turn that wiring into reusable infrastructure. Still, the coverage makes clear that connecting agents is only the first step toward agents that can be trusted to work together.

How the stack fits together

Coverage is close to unanimous that the two main protocols complement each other. MCP handles the vertical link between one agent and the systems it calls. A2A handles horizontal coordination between peer agents, which may run on different frameworks from different vendors.2228 The Linux Foundation describes the same split: A2A covers agents coordinating across organizational boundaries, and MCP covers agents reaching internal tools and data.21 One practitioner guide sums up the pattern in practice. An agent exposes an A2A endpoint so others can delegate to it, then uses MCP internally to reach the tools it needs for the job.14

Governance shifted over the past year. In December 2025, Anthropic donated MCP to the Agentic AI Foundation, a directed fund under the Linux Foundation co-founded by Anthropic, Block and OpenAI, with backing from Google, Microsoft, AWS, Cloudflare and Bloomberg.20 Google had already given A2A to the Linux Foundation in June 2025.30 One report says Google formally moved A2A into the Agentic AI Foundation on August 20, 2026, putting both protocols under one governance structure.28 That account also says the foundation has grown to more than 250 member organizations. However, its claim that the foundation started in March 2026 with 49 founding members does not match other reporting that dates the foundation to December 2025.2820 The exact timeline is unclear, but every account agrees that both protocols are now governed by a neutral body rather than by a single vendor.

A2A's first year: big numbers, early days

A2A has moved faster than many expected. At its one-year mark in April 2026, the Linux Foundation said more than 150 organizations supported the protocol. Production deployments were running in supply chain, financial services, insurance and IT operations.21 Version 1.0, the first stable specification, added multi-tenancy, updated security flows, and Signed Agent Cards that let agents prove their identity cryptographically.21 Microsoft built A2A into Azure AI Foundry and Copilot Studio, and AWS added support in Amazon Bedrock AgentCore Runtime.21 The SDKs grew from Python alone to five languages, and the main repository passed 22,000 GitHub stars.21

Observers outside the foundation are more cautious. One commentator says A2A is less proven in production than MCP, even though its governance and adoption are taking agent-to-agent communication beyond any single platform.20 Another says the real test is not who announces support. It is who keeps A2A running after the first serious operational incident.25 A guide from an API gateway vendor says the core protocol is stable but that architects should still expect to build workarounds for gaps such as token downscoping and registry standardization.27 The foundation's own roadmap lists an interoperability specification, registry consolidation and security best practices as unfinished work.21 Read together, the picture is that A2A's specification is ready, but the operating practices around it are not yet.

MCP: widespread, but adoption numbers disagree

MCP is the more mature of the two, but its adoption figures vary widely. The protocol's maintainers reported more than 97 million monthly SDK downloads and over 10,000 active servers in December 2025. They also listed client support in ChatGPT, Claude, Cursor, Gemini, Microsoft Copilot and Visual Studio Code.20 Enterprise production figures are less consistent. Several outlets say 78% of enterprise AI teams have MCP-backed agents in production.45 One analysis that rechecked that number swapped it for a Stacklok survey finding that 41% of software organizations were in limited or broad production, with only 12% in broad production.2 The same review counted 9,652 current records in the official MCP registry in May 2026.2 One later report says a follow-up Stacklok survey did find the 78% figure.9 The more conservative 41% number has the clearest sourcing, so it is the safer basis for planning.

The specification itself is also changing. In July, MCP's Enterprise-Managed Authorization extension became stable. It moves access decisions to a company's identity provider and replaces repeated consent prompts on each server with a single sign-in. Anthropic, Microsoft and Okta have adopted it, along with servers from Asana, Atlassian, Figma and Linear.1 On July 28, maintainers finalized a major revision that removes session tracking at the protocol level, making MCP stateless at that layer.7 Both changes target the same problem: running MCP at enterprise scale.

Interoperability is reaching products

The protocols are now showing up in products people use every day. At WebexOne on October 7, Cisco presented a rebuilt Webex where collaborative agents can join spaces, meetings and calls. Anthropic's Claude, running on Claude Managed Agents, can pick up shared team context.38 Cisco also added an MCP integration with OpenAI's dots, which places two rival model providers inside one suite.38 TechTarget reported that Cisco separately announced plans to support both A2A and MCP.36 Most of these features are not due until early 2027, and Cisco has not given dates.38

Similar efforts are under way across the industry. JetBrains is promoting its Agent Client Protocol as a shared layer between coding agents and IDEs, comparing it to what the Language Server Protocol did for language features.11 In telecom, TM Forum's A2A-T variant is built to coordinate agents across billing, service and network domains.17 In payments, Ant International, Mastercard and Visa have started work on an interoperable Know-Your-Agent framework. It builds on each company's own agent protocol so that trust signals carry across card and wallet networks.18 In each case, organizations are adapting the general idea of interoperability to their own sector.

Making agents talk is the easy part

The most important point across the coverage is that shared message formats do not create shared understanding. An O'Reilly analysis notes that a schema describes the shape of a tool's arguments, while its meaning is left to free-text descriptions that a model has to interpret.20 The telecom analysis argues that without a shared semantic layer, agents can be technically connected while the workflows they support still fail.17 Reliability data points the same way. A Berkeley study of multi-agent systems found failure rates between 41% and 86.7%, and traced most failures to agents misunderstanding one another rather than to any single model being wrong.16

Security is the other unresolved issue. In May 2026, the U.S. National Security Agency said MCP's rapid spread had outpaced its security model.7 Independent scans cited in one report found exploitable flaws in 30% to 82% of public MCP servers, depending on the scan.9 A Gravitee survey found that only 14.4% of organizations had fully security-cleared all their deployed agents.6 In health care, an Imprivata survey found that 72% of industry leaders said agents were being deployed without formal IT approval.31 Washington is also taking a closer interest. NIST launched an AI Agent Standards Initiative in February focused on interoperability and security.15 This month, following incidents disclosed by Anthropic, the White House said AI companies must report and fix security incidents involving their models.34

The takeaway

In my reading, the coverage supports a clear conclusion. The protocol war is effectively over, and shared standards won. Anthropic, OpenAI, Google, Microsoft and AWS now compete on models and applications while jointly governing the interfaces between them.2028 For enterprises, that removes most of the risk of betting on one vendor's protocol.

The remaining work is less visible and harder: semantics, identity, authorization at each handoff, and auditing chains of delegated tasks. One playbook argues that a network of interoperable agents needs more governance than a single agent, because an unaudited chain of handoffs can put the whole organization at risk.14 Companies that treat A2A and MCP as only a connectivity upgrade are likely to run into those problems. Companies that also invest in registries, scoped permissions and observability will be in a better position to benefit from agents that can work together.