Open Standards Instead of Going It Alone: What A2A Joining the Foundation Means for Your Agent Investment

Sometimes a piece of news looks at first glance like a detail for specialists — and turns out on closer inspection to be solid information for every purchasing decision. On August 17, 2026, Google's Agent2Agent protocol (A2A) officially became a "hosted project" of the Agentic AI Foundation (AAIF), the neutral umbrella organization under the Linux Foundation that has already been home to Anthropic's Model Context Protocol (MCP) since December 2025. Two of the most important technical standards for AI agents now sit under the same vendor-neutral roof for the first time.

For IT decision-makers in the mid-market, that sounds at first like a footnote from the world of standardization. It isn't. Anyone investing in agent software today — whether for customer service, sales, accounting, or production planning — is often binding themselves to an architecture for years. Whether that architecture rests on open, widely supported standards or on one vendor's proprietary solo run helps determine how expensive a later switch becomes, how well your agent gets along with the systems of your partners, suppliers, and customers — and how dependent you remain on the whims of a single manufacturer.

This article explains what actually happened on August 17, 2026, why MCP and A2A solve two different but complementary problems, and — this is the part that actually matters — which questions you should ask yourself before signing the next line of an agent contract.

What actually happened on August 17, 2026

A look back helps put this in context. A2A was by no means new: Google had already handed the protocol to the Linux Foundation in June 2025, with AWS, Cisco, Google, Microsoft, Salesforce, SAP, and ServiceNow as founding members. At the time, it was a standalone project within the broader Linux Foundation structure.

In parallel, Anthropic, Block, and OpenAI founded the Agentic AI Foundation on December 9, 2025 — backed by Google, Microsoft, AWS, Cloudflare, and Bloomberg — as a specialized home for agent standards. Founding projects were MCP (Anthropic), the open-source agent framework goose (Block), and AGENTS.md (OpenAI), a standard format for instructions to coding agents.

The August 17, 2026 move now brings A2A into this same specialized AAIF structure, directly alongside MCP. Important for anyone already working with A2A: this is a pure governance change. The specification version stays at 1.0.0, and existing Agent Cards and SDKs keep working unchanged. Nobody needs to rewrite code because of this news.

What does change is the signal it sends: the two currently most influential interoperability standards for AI agents now run under the same organizational roof, with the same governance rules, the same membership model, and — at least formally — the same vendor neutrality.

MCP and A2A: two standards, two layers

The two protocols are often mentioned in the same breath in coverage, but they solve different problems. It's worth keeping them cleanly separated, because you should ask both questions separately for every agent solution.

MCP standardizes the vertical connection: How does a single agent access tools, databases, APIs, and file systems? MCP is the interface between an agent and its toolbox — your CRM, your accounting software, your file server.

A2A standardizes horizontal communication: How does one agent hand a task to another agent, and how does it get the result back? This becomes relevant as soon as agents from different vendors or frameworks — yours, a supplier's, a customer's — need to work together across organizational boundaries.

An example makes the difference tangible: a purchasing agent at your company uses MCP to access your inventory management system and your invoice database. But if that agent needs to trigger an order with a supplier whose sales agent runs on a completely different technical foundation, A2A comes into play — it's the shared language the two agents can even use to communicate at all, without anyone having to code a proprietary direct integration.

The fact that both standards now sit under the same roof changes nothing about this division of labor. It merely means the likelihood increases that vendors will support both layers consistently and from a single source, instead of offering a different, possibly incompatible solution for each layer.

Why this matters for your agent investment

The raw numbers show how quickly these standards have taken hold. The AAIF counted around 40 to 49 founding members in December 2025; by mid/late August 2026 it reportedly had more than 250 members, with platinum members including AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft, and OpenAI. According to a Linux Foundation press release from April 2026, A2A is supported by more than 150 organizations — up from a good 50 in April 2025. The core repository logged more than 22,000 GitHub stars at that point, with production-ready SDKs for five languages: Python, JavaScript, Java, Go, and .NET. MCP, in turn, counted more than 10,000 active public servers at the time of the AAIF's founding in December 2025 — a number that has presumably grown further since, though no more recent, reliable figure is available.

Important: these two sets of figures — "250+ AAIF members" and "150+ A2A supporters" — come from different reports at different points in time and count slightly different things. They shouldn't be used interchangeably, but both point the same direction: an enormous and still-growing ecosystem of vendors has committed to these two standards.

For your investment, this means in practice: A2A is now natively supported on major cloud platforms, including Microsoft Azure AI Foundry and Copilot Studio, AWS Bedrock AgentCore Runtime, and Google Cloud. Reports indicate that, among others, Huawei's assistant "Celia" on HarmonyOS and Tencent WeChat already use A2A in practice for cross-platform agent coordination. If you invest in an agent solution built on these standards today, you're not buying into a niche product, but into infrastructure carried in parallel by the world's largest cloud providers.

This is the real value for you as a mid-market company: a vendor switch, an integration with a new partner, or a migration to a different cloud environment becomes more predictable when the underlying protocols are openly documented and broadly supported — instead of being tied to a single, closed API format from one vendor.

Why this isn't an academic argument becomes clear when you look at the rest of the market. According to Flexera's "State of the Cloud Report 2026," 89 percent of companies already run two or more cloud providers in parallel — a multi-vendor reality that carries over into the agent world as soon as you work with partners, suppliers, or customers running different platforms than you do. At the same time, Gartner forecasts that more than 40 percent of agentic AI projects will be abandoned by 2027, while around 40 percent of enterprise applications are expected to contain task-specific AI agents by the end of 2026. Taken together, these two figures paint a clear picture: agent adoption is growing fast, but a significant share of projects fail — often not because of the AI's performance itself, but because of poor integration capability, missing governance, or exactly the kind of silent vendor lock-in that open standards are meant to prevent. Anyone basing their agent investment on an architecture demonstrably carried by a broad ecosystem reduces a real risk of failure — not just a theoretical one.

The reality check: what "open standard" doesn't guarantee

This is where things get more nuanced, and this is exactly the part that deserves your attention before you base a purchasing decision solely on the label "A2A-compatible" or "AAIF member."

First: formal neutrality doesn't automatically mean a broad diversity of contributors. An independent technical assessment from the analysis firm Pebblous notes that, despite the formally vendor-neutral governance, three organizations account for more than 51 percent of documented contributions to the AAIF projects. This figure comes from a single blog analysis, not from an official Foundation statistic — it should therefore be read as an assessment, not a confirmed fact. But it's a useful pointer: "open standard" doesn't necessarily mean "carried evenly by many."

Second, and this is the technically more important point: signed Agent Cards in A2A verify an agent's identity — not its permissions. The actual rights and policy model, i.e., the question of who may talk to whom about what and exchange what data, is deliberately left open in the A2A specification. This is exactly why Google Cloud and PayPal separately built the Agent Payments Protocol (AP2), now supported by more than 60 organizations, which governs payment authorization between agents where A2A itself provides no answer.

Downstream agents in an A2A chain frequently treat the output of the preceding agent as one hundred percent trustworthy — instead of treating it as a claim to be verified. That is exactly where the risk of error propagation in multi-agent systems lies.

This observation from the cited technical analysis doesn't concern the protocol specification itself, but the way vendors implement A2A in practice. For you as a buyer, this means: the question "does the vendor support A2A?" is necessary but not sufficient. The question that actually matters is how the vendor concretely handles authorization and chains of trust — a layer the open standard deliberately does not prescribe.

Third, a point for context, not for dramatization: the AAIF board is led, among others, by David Nalley (AWS), and the technical committee, according to the same source, by David Soria Parra (Anthropic, co-developer of MCP). This shows that "neutral" in the sense of governance structure isn't synonymous with "independent of the large US hyperscalers." For most mid-market companies, that's not a disqualifying factor — but anyone who, for data protection or sovereignty reasons, deliberately values independence from US cloud providers should keep this point on their radar rather than overweighting it.

The purchasing-criteria checklist for your business

From all this, a practical checklist can be derived that you should run through for every agent offer — whether it's standard software, a custom-built solution, or an offer from a consulting partner.

The MCP question: Can I connect my own tools, databases, and APIs without permanently binding myself technically to this one vendor? Is MCP natively supported and openly documented, or is there only a proprietary analog that looks similar but isn't compatible?

The A2A question: Can my agent actually work with agents from other vendors, partners, or customers — or does collaboration stay confined to a single manufacturer's closed ecosystem?

The authorization question: How does the vendor concretely verify which data and rights an external partner agent actually receives via A2A? Is there documented verification of chains of trust, or is the output of upstream agents adopted unchecked?

The portability question: Can your agent configurations be exported? Can you switch the underlying language model without rebuilding the entire automation? Standard membership alone — i.e., a vendor's membership in the AAIF or the label "A2A-compatible" — is no guarantee against de facto lock-in if exportability and model portability are missing in practice.

The continuity question: What happens to your investment if a standard's governance changes again, as happened on August 17, 2026? In the current case, there were no breaking changes — the specification stayed at version 1.0.0. That's a good sign for the ecosystem's maturity, but not a free pass to stop asking this question going forward.

These five questions can't be answered by looking at a product website. They require a conversation with the vendor — ideally one in which concrete technical evidence, not marketing promises, is put on the table.

What this means for the coming months

The consolidation of MCP and A2A under the roof of the Agentic AI Foundation is, in the best sense, unspectacular: it doesn't change a single line of code, it forces no one to migrate, and it poses no immediate threat to existing investments. That's exactly what makes it significant, though. Standardization in the software industry rarely happens through a single dramatic cutover, but through exactly these kinds of gradual consolidations, where more and more vendors, frameworks, and cloud platforms agree on the same technical foundation.

For the mid-market, this means: the point at which "supports MCP and A2A" becomes a baseline requirement instead of a differentiator is getting closer. Anyone still investing in closed, proprietary agent architectures today should be aware that they're going against a trend carried simultaneously by practically all major cloud and AI vendors — from Anthropic to Google to Microsoft and AWS.

At the same time: a standard alone doesn't make good governance. The authorization questions A2A leaves open, the uneven distribution of contributions, and the continuing closeness of the governance leadership to the major hyperscalers are real points you should factor into your own risk assessment — regardless of how neutral a foundation describes itself.

Anyone building an agent strategy for their company should keep both layers in view: the technical question of whether a vendor actually uses open standards, and the practical question of how they concretely implement security, authorization, and portability. One doesn't replace the other.