MCP, A2A, and the agent interoperability stack

Last verified: 10 November 2026

The distinction that clears up most confusion

MCP connects an agent to tools and data. Your CRM, your product catalog, your internal search, your ticketing system. One agent, many capabilities. Inside your walls.

A2A connects an agent to another agent. Your procurement agent talking to a supplier’s fulfillment agent. Two autonomous systems, different owners, negotiating and handing off work. Across organizational boundaries.

If you’re wiring Claude into your data warehouse, that’s MCP. If you’re letting your logistics agent coordinate with a carrier’s agent, that’s A2A. Most organizations need the first one long before they need the second.

They’re designed to compose. A2A carries the conversation between agents; each agent uses MCP to reach its own tools. Nobody has to choose. For where these sit relative to commerce and payments, see the agentic protocol landscape.

Model Context Protocol

It’s not Anthropic’s anymore

Anthropic created MCP and released it in late 2024. On 9 December 2025 it donated the protocol to the Agentic AI Foundation, a directed fund under the Linux Foundation that Anthropic co-founded with Block and OpenAI, with support from Google, Microsoft, AWS, Cloudflare, and Bloomberg.

The founding projects were MCP, Block’s goose, and OpenAI’s AGENTS.md. Agentgateway joined on 4 June 2026, Envoy AI Gateway followed as Agent Router, and A2A moved in on 17 August (see below). On 22 October four more arrived — Agent Skills (from Anthropic), Open Responses (OpenAI), OpenEnv (Hugging Face) and OpenSandbox (Alibaba) — bringing the foundation to ten hosted projects. Membership passed 240 organizations in August.

Governance is worth understanding because of how it’s structured. Authority runs through individuals, not company seats — Lead Maintainers, then Core Maintainers, then Maintainers. No organization holds a bloc. The current Core Maintainer roster spans Anthropic, Microsoft, Google Cloud, and AWS, and every one of them holds their seat personally.

Specification changes go through a proposal process (SEPs) that moved to a pull-request workflow in late 2025. There’s a published feature lifecycle with a twelve-month minimum deprecation window. This is normal open-source governance, and its arrival is the strongest signal that MCP is infrastructure rather than a vendor feature.

The July 2026 release broke things on purpose

The current specification is 2026-07-28. The previous revision was 2025-11-25. The gap between them is large, and if you have MCP servers in production you need to know what’s in it.

The core went stateless. The initialize handshake is gone. Protocol version, client capabilities, and client info now travel per-request in _meta alongside an MCP-Protocol-Version header. A new server/discover RPC handles up-front negotiation when you want it.

Sessions were removed. Mcp-Session-Id is out of Streamable HTTP. Cross-call state now moves through explicit server-minted handles passed as tool arguments.

Server-initiated requests were replaced. Instead of a server calling back to the client, a server returns an input_required result carrying its requests, and the client retries with responses attached. This is called Multi Round-Trip Requests.

Sampling, Roots, and Logging are deprecated. Twelve-month window. Migration paths: tool parameters instead of Roots, direct LLM provider APIs instead of Sampling, stderr or OpenTelemetry instead of Logging.

Tasks moved out of core into an official extension.

Authorization tightened. Clients must validate an RFC 9207 iss parameter when one is present, credentials are bound to the authorization server that issued them, and Dynamic Client Registration is deprecated in favor of Client ID Metadata Documents (it still works for servers that lack them; no removal date yet).

Transport housekeeping. Streamable HTTP requests now carry Mcp-Method and Mcp-Name headers, SSE stream resumability is gone (a broken stream loses the request; the client re-issues it), list results carry ttlMs and cacheScope cache hints, and the legacy HTTP+SSE transport is formally deprecated.

The through-line is that MCP is being rebuilt to scale like ordinary web infrastructure. Stateless requests survive load balancers, retries, and cold starts in a way stateful sessions never did. It’s the right call, and it costs work.

What to do about it

If you shipped MCP servers before mid-2026, you’re on a spec revision with deprecated features. Nothing breaks tomorrow — the twelve-month window is real, and removal can come no earlier than July 2027 — but “we’ve deployed MCP servers” is no longer a statement about whether they’ll work next year.

Three questions worth answering internally:

• Which spec revision does each server target? If nobody knows, that’s the finding.
• Do any of them depend on Sampling, Roots, or Logging? Those are on the clock.
• Do any depend on session state? That’s the migration with real work in it.
• Do your gateways, proxies or rate limiters route on session IDs or SSE resumption? They need to read the new Mcp-Method and Mcp-Name headers instead.

The spec itself hasn’t moved since. 2026-07-28 is still the current revision; the working draft holds only small fixes, and the maintainers’ 22 August roadmap names priorities (server-initiated events, agent identity and delegation, HTTP-native transport) but no date for the next release.

Where MCP actually is

The four Tier 1 SDKs (TypeScript, Python, Go, C#) all speak 2026-07-28, with Rust in beta. AWS shipped a generally available MCP server for all AWS services in May 2026. Google announced more than fifty managed MCP servers in April.

One thing that has not happened: the MCP Registry has not reached general availability. It launched in preview in September 2025 and its API froze at v0.1 in October. The documentation still warns about breaking changes and data resets. Don’t build a hard dependency on it yet.

Agent2Agent

The stability milestone most people missed

Google announced A2A in April 2025 and donated it to the Linux Foundation on 23 June 2025. Founding partners were AWS, Cisco, Google, Microsoft, Salesforce, SAP, and ServiceNow. IBM’s separate Agent Communication Protocol merged into it in August 2025 and is no longer a standalone spec.

On 17 August 2026 A2A became a hosted project of the Agentic AI Foundation, alongside MCP. It keeps its own Technical Steering Committee (AWS, Cisco, Google, IBM Research, Microsoft, Salesforce, SAP, ServiceNow). So the two protocols now share a neutral home, which was not true when this page was first written.

Through 2025 it moved fast and broke things across 0.x releases. The reasonable position then was to watch it.

That position expired on 12 March 2026, when A2A shipped v1.0.0 — its first stable release. It carried breaking changes to the interaction protocol, added multi-tenancy and signed Agent Cards, and moved to a stateless, layered, web-aligned architecture with three bindings: JSON over HTTP, gRPC, and JSON-RPC. Version 1.0.1 followed on 28 May with bug fixes; no later release has shipped.

At the one-year mark A2A had 150+ supporting organizations, 22,000+ GitHub stars, and five SDKs. It’s live in Azure AI Foundry, Microsoft Copilot Studio, and Amazon Bedrock AgentCore Runtime, with reported production use in supply chain, financial services, insurance, and IT operations.

The A2Family

Google now groups four things together, and the grouping is useful even if you’re skeptical of the branding:

• A2A — agents talking to agents.
• AP2 — agents paying (an extension of A2A, now donated to the FIDO Alliance).
• A2UI — agents rendering interfaces to humans.
• UCP — agents transacting commerce (which can run over A2A or MCP).

That’s a coherent stack. It’s also mostly Google-originated, which is worth naming — though AP2 went to FIDO and A2A went to the Linux Foundation, so the parts that matter most are now stewarded elsewhere. The commerce corner of it is covered in Agentic commerce protocols, explained, and the payments corner in Agent payment rails, explained.

AGNTCY: the layer underneath

AGNTCY handles the infrastructure problems that show up once you have more than a handful of agents. A federated Agent Directory for discovery. Identity with verifiable credentials and policy-based access. SLIM for secure low-latency messaging. OASF, a schema framework. Plus observability and evaluation tooling.

Cisco open-sourced it in March 2025 and it became a Linux Foundation project on 29 July 2025, with formative members Cisco, Dell Technologies, Google Cloud, Oracle, and Red Hat and 65+ supporting companies.

It’s explicitly complementary: A2A agents and MCP servers are discoverable through AGNTCY directories, and AGNTCY’s identity work covers MCP servers directly. It sits beneath both rather than alongside them.

On 10 November 2026 the Linux Foundation welcomed the Agent Name Service (ANS), a DNS-based identity and discovery standard, as a project of its own; Cisco’s Outshift says ANS is being integrated into the AGNTCY Directory, and the Directory already supports verifying ANS identities. That is the first clear 2026 milestone for the AGNTCY ecosystem. We’ll still be straight about the rest: the code is actively maintained, and the directory added automated security scanning for MCP servers and agent skills, but we found no published component versions and no named production deployments. That’s consistent with a technically healthy project going through a quiet stretch. It’s also consistent with a project losing momentum. We can’t tell from outside, and we’d rather say that than guess.

Note also that AGNTCY is not part of the Agentic AI Foundation, despite the similar-sounding “AGNTCon” event name. Cisco and Oracle participate in AAIF separately as Gold members.

The governance map, because it explains a lot

Where these standards live now tells you more about their future than their feature lists do.

• Linux Foundation — Agentic AI Foundation (ten projects): MCP, A2A (since August 2026), goose, AGENTS.md, agentgateway, Agent Router, Agent Skills, Open Responses, OpenEnv, OpenSandbox.
• Linux Foundation — standalone projects: AGNTCY, Agent Name Service, x402 Foundation.
• FIDO Alliance: AP2, plus the Agentic Authentication and Payments working groups.
• W3C Web Machine Learning Community Group: WebMCP (incubation, not standards track).
• Vendor-controlled: ACP (OpenAI, Stripe, Meta on a TSC — no foundation), Visa TAP (Visa).
• Industry consortium: UCP (a Tech Council of Google, Shopify, Etsy, Target, Wayfair, Amazon, Meta, Microsoft, Salesforce, Stripe).

Two useful reads from that list. First, the foundation column got long in 2026 — that’s the market voting for neutrality. Second, the vendor-controlled column still contains real, widely-deployed protocols, and there’s nothing wrong with that as long as you price the risk consciously.

What we’d actually do

Start with MCP, and audit what you already have. It’s the most settled standard in agentic AI and the one most likely to already be in your stack. The July 2026 spec change is the more urgent item than any new adoption.

Treat A2A as ready. Stability shipped in March, and it now sits beside MCP at the Agentic AI Foundation. If agent-to-agent workflows are on your roadmap, the technical reason to wait is gone. The remaining reasons are organizational — do you actually have a counterparty ready to talk to?

Watch AGNTCY, don’t bet on it. The problems it solves are real and arrive later. Nothing about waiting costs you.

Don’t build a house on the MCP Registry. Still preview, still v0.1.

Check your gateways. Anything that routes MCP traffic by session or relies on SSE resume needs the new headers and retry behavior before you move servers to 2026-07-28.

Ask vendors for dates, not names. “We support MCP” means nothing on its own. “We support MCP 2026-07-28, and here’s our deprecation plan for Sampling” means something.

Common questions

What is the difference between MCP and A2A?

MCP connects an agent to your tools and data. A2A connects your agent to someone else’s agent. They are complementary rather than competing — A2A handles agent-to-agent across organizational boundaries, MCP handles agent-to-tool inside them. That distinction clears up most of the confusion about the two.

Is MCP still controlled by Anthropic?

No. Anthropic created MCP and donated it in December 2025, co-founding the Agentic AI Foundation — a directed fund under the Linux Foundation — alongside Block and OpenAI. Governance runs through individual maintainers rather than company seats, so no vendor holds a controlling stake. A2A joined the same foundation in August 2026.

What did the July 2026 MCP spec change?

It broke things on purpose. The initialize handshake is gone and the core is now stateless, with version and capability negotiation carried per-request in _meta plus the MCP-Protocol-Version header. Protocol-level sessions and Mcp-Session-Id were removed from Streamable HTTP. Server-initiated requests were replaced by Multi Round-Trip Requests. Sampling, Roots, and Logging are deprecated on a twelve-month window, Tasks moved out of core into an extension, and Dynamic Client Registration is deprecated in favour of Client ID Metadata Documents.

Should I build on A2A yet?

Yes, and the reason changed this year. A2A reached version 1.0 in March 2026, its first stable release, with v1.0.1 following in May, and it moved under the Agentic AI Foundation in August. The wait-and-see position that made sense in 2025 does not anymore. It has 150+ supporting organizations and runs in Azure AI Foundry, Microsoft Copilot Studio, and Amazon Bedrock AgentCore Runtime.

Can I rely on the MCP registry?

Not yet. The registry is still in preview — the API froze at v0.1 in October 2025 and has not reached general availability. Do not build a dependency on registry stability.

Sources

MCP specification 2026-07-28 changelog · MCP governance · Anthropic donates MCP, AAIF formed · Agentic AI Foundation · MCP Registry · A2A joins AAIF · AAIF adds four projects · MCP roadmap · MCP 2026-07-28 release · Agent Name Service joins the Linux Foundation · Google donates A2A to the Linux Foundation · A2A releases · A2A passes 150 organizations · Meet the A2Family · Linux Foundation welcomes AGNTCY · AGNTCY docs

Want your MCP surface audited?

We run MCP and A2A in production on our own platform — including the spec migration this guide describes — on top of Neural Core and the marketing infrastructure underneath it. If you need to know which of your servers survive the next twelve months, that’s a conversation worth having.

Talk to us More resources