On June 17, 2026, Google’s developer blog published the specification for Agentic Resource Discovery (ARD). Ten more companies signed on as backers the same day: Microsoft, Salesforce, ServiceNow, Snowflake, Cisco, Databricks, GitHub, GoDaddy, Hugging Face, and NVIDIA. One name is conspicuously missing if you work with AI daily: Anthropic is not there. Neither is OpenAI.

That is not an oversight. It is the thesis of this post: the fight that actually matters in 2026 is not which model reasons best. It is who controls the protocol an AI agent uses to find and connect to your systems. Anthropic has spent eighteen months owning that layer through the Model Context Protocol, the de facto standard for wiring AI assistants to tools and data. Five of the largest enterprise software companies on earth just announced they are building their own discovery layer, separately, and did not invite the two labs that build the models that layer will end up serving.

The protocol, not the model, is where this fight lives

I wrote about the harness in May and about governance in July: the layer wrapped around the model is where competitive advantage actually lives in 2026, not the model itself. ARD is the clearest confirmation of that thesis I have seen yet, and it is coming from an unexpected direction. No consultant wrote this. It is being built, with real budget and coordination between direct competitors, by the five companies with the most to lose if Anthropic and OpenAI end up controlling how every agent walks into their systems.

Look at it from Salesforce’s side, or ServiceNow’s. Each one sits on decades of customer data living inside its own platform. If the only standard path for an AI agent to discover and use that data runs through a protocol Anthropic designs, governs, and defines, then Anthropic, not Salesforce, not ServiceNow, ends up writing the access rules to that data. That is exactly what an independent analysis from softmaxdata named without hedging: “a window into the central power struggle of this era of AI, between the companies that supply the capabilities and the companies that want to own the interface.” No executive has to say this out loud in a press release for it to be true. You just have to look at who signed and who did not.

What ARD actually does

ARD does not replace MCP. It solves a different problem, one step earlier: how does an agent find the right tool among hundreds of thousands scattered across domains no single company controls. The spec defines two pieces. First, an ai-catalog.json manifest each organization publishes at a well-known path on its own domain, describing what capabilities it offers. Second, a registry API that crawls and indexes those catalogs and answers natural-language discovery queries with ranked matches.

With that, ARD answers three questions an agent cannot currently answer on its own: where does the right capability live, which one should I use among similar ones, and how do I verify it is safe to connect to. Once the agent knows which tool to use, it invokes it over its native protocol, which can perfectly well be MCP, Google’s own A2A, or a plain API. Google and Microsoft’s official documentation is explicit here: ARD “is not a replacement for MCP, A2A, or Skills.” It is a discovery layer, not an execution layer. Apache 2.0-licensed, governed through a Linux Foundation working group, currently at spec version 0.9 (draft). GitHub already shipped Agent Finder for Copilot, and Hugging Face shipped its own discovery tool the same day the spec went live: real implementations, not just a document.

What the press releases say, and what the coverage says

There is a nuance here worth keeping sharp, not flattening into one story. Read Google’s announcement or Microsoft’s, and the tone is carefully neutral. Google says “there is no standard way to answer those questions across organizations.” Microsoft compares ARD to what a search engine did for web pages, and adds: “the goal isn’t a single global catalog of every resource. There will be many discovery services.” Neither official post mentions Anthropic. Neither mentions OpenAI. This is industry-coordination language, not confrontation language.

Then there is the coverage. The Information’s headline on the story was “Google and Microsoft Team Up to Beat Back Anthropic and OpenAI.” CIO framed its piece around what enterprise IT teams need to solve while no unified catalog exists yet. Both readings are true at once. The companies backing ARD are never going to say in a press release that they built this specifically to avoid depending on Anthropic. That would break the “open standard for everyone” story they need for third parties to adopt it. But the exact roster of who signed, and more importantly who did not, says what the press release never will.

Why AI Maestro still runs on Claude and MCP

A post like this is easy to misread as a signal to hedge your bets. That is not the read. AI Maestro, IQ Source’s own service, runs on Anthropic’s Claude and on MCP to connect agents to a client’s systems. That decision predates ARD, and it still holds today for a simple reason: MCP has eighteen months of real adoption, thousands of community-built public servers, and native support across every frontier lab, Anthropic included, but also Google DeepMind and Microsoft. ARD is a 0.9 draft with reference implementations that just shipped. One is proven infrastructure. The other is a coordination bet between competitors that still has to prove eleven large companies can keep a shared standard aligned without each one pushing its own variant.

What this post does change is the context that decision gets made in. We did not choose MCP because it was the only option on the table, or because we did not know enterprise platforms would eventually want control over this layer. We chose MCP because, so far, it is where real adoption lives, and real adoption is what makes an agent connected to your ERP or your CRM actually work the week you ship it, not the week the standard matures. If ARD gains genuine traction over the next twelve months (published catalogs, not just press releases), it becomes an additional layer sitting on top of MCP, not a replacement that forces a rewrite. The two do not really compete in practice. They compete for who governs the conversation about enterprise agents in the boardroom.

What a CTO should watch, not what they should announce

There is no urgent action to take on ARD this week. The spec is a month old, real adoption is still thin, and no client is going to ask you tomorrow to publish an ai-catalog.json. What is worth doing is asking a different question than the one most IT teams are asking. When you evaluate your next AI agent vendor, or your next integration, ask what protocol it speaks, not just what model runs underneath. The model you use this quarter you will swap out next quarter. The protocol you pick to connect your systems is the one still standing when nobody remembers which model version was trendy in July 2026.

If you cannot answer that question clearly for your own company yet, it is exactly the kind of map we build in the discovery phase of AI Maestro: identify which layer your real dependency lives in before you commit to a standard that may or may not still be alive in two years. I already covered why the operating harness is the moat and why the governance control tower is one too. ARD adds a third layer to that same list: the connection protocol. None of the three come pre-packaged. All three have to be decided with your own judgment, and that decision outweighs which model you picked this quarter.