Five Giants Move to Break Anthropic's Grip on the Agent Stack

Protocol wars sound abstract until you remember the last two decided who owned the internet. TCP/IP settled how machines find each other; HTTP settled how they talk. Both looked like plumbing arguments at the time. Both turned into the substrate every business on earth now rents. On July 13, five of the largest names in enterprise software moved to make sure the next one doesn't get decided by Anthropic.

Google, Microsoft, Salesforce, Snowflake, and ServiceNow agreed to jointly support a shared backend protocol for AI agents — the connective layer that lets an autonomous agent reach into a company's data, tools, and workflows. The framing was not subtle. The alliance is a deliberate counter to Anthropic and OpenAI, whose standards have quietly become the default over the past eighteen months. The five aren't building a better model. They're contesting the wiring underneath the models, which is the part that's genuinely hard to rip out once it's installed.

Why these five, and why now

Look at what this group actually controls. Salesforce, Snowflake, and ServiceNow sit on top of most of the world's enterprise data and business processes — CRM records, the data warehouse, the IT and HR workflow engine. Google and Microsoft own the clouds all of it runs on. A protocol co-signed by those five isn't a startup's hopeful spec sheet. It's a standard that arrives pre-installed in the systems where corporate work already happens.

That matters because of how the incumbent won. Anthropic's Model Context Protocol (MCP) didn't take the enterprise by decree. It spread the way good developer standards spread — it was open, it was simple, and it solved the exact problem everyone had at once: how do you let a model call a tool or read a system without hand-writing a bespoke integration for every pairing. Eighteen months of that quiet adoption and MCP became the thing you build against by default. Default status is the whole prize. Whoever owns the agent-connection standard gets first position in every enterprise AI deployment for the next decade, the same way HTTP got first position in every website.

The five giants watched a competitor accumulate that position and decided they'd rather not build their agent stacks on a rival's foundation. So they're pooling their distribution to offer enterprises a credible alternative — one where the standard is governed by a consortium that includes them, not by the company they're trying to out-sell.

The part where everyone is fighting and cooperating at once

Here's the twist that makes this more than a normal platform brawl. All five of these companies — plus Anthropic and OpenAI — simultaneously participate in the Linux Foundation's Agentic AI Foundation, the neutral body set up to build shared, open standards for exactly this layer. Everyone is cooperating inside the foundation and knife-fighting in the market at the same time.

That's not hypocrisy so much as the standard playbook for how open standards actually get set. You contribute to the neutral spec so no single vendor can be accused of owning it, and you race to ship the best real-world implementation so that your version becomes the reference everyone copies. Google has already run this move once. Its Agent2Agent (A2A) protocol — the standard for how agents talk to other agents, as opposed to MCP's focus on agents talking to tools — launched with more than fifty partners, has now reached roughly 150 organizations running it in production, and hit version 1.2 under Linux Foundation governance. Donating A2A to the foundation didn't weaken Google's hand. It laundered a Google project into an "open standard" that Google still leads.

The five-company alliance is that same strategy applied to the backend layer MCP currently owns. Neutral governance on paper, competitive positioning in practice.

What actually gets decided here

Strip away the diplomacy and the stakes are concrete. An enterprise standardizing its agent infrastructure this year is choosing a foundation it will be stuck with for a long time — migrations at this layer are brutal, because every tool integration, every permission boundary, every audit trail gets rebuilt if you switch. That lock-in is precisely why the position is worth fighting for, and precisely why the incumbents' rivals can't afford to let MCP calcify into the only option.

For the rest of the market, the immediate effect is fragmentation before consolidation. In the near term, builders face a messier map: MCP for tool access, A2A for agent-to-agent coordination, and now a coalition-backed backend protocol angling to sit under all of it. The protocol stack is layering up faster than any one team can cleanly adopt. Expect the winners to be the frameworks and gateways that abstract the whole mess — the ones that let a developer target "the agent layer" without betting the company on which specific standard survives.

The deeper signal is about where the industry now thinks the money is. A year ago the fight was over who had the smartest model. The best models are now close enough that no one can win on raw capability alone, so the giants have moved the battle down a layer — to the connective tissue that determines whose infrastructure every agent runs through. That's a tell. When the incumbents stop competing on the product and start competing on the protocol, it means they've concluded the product is a commodity and the plumbing is the moat.

Anthropic built that moat first, in the open, by being early and useful. Now the five companies with the most enterprise distribution on the planet have decided to build a competing one — inside the same foundation, under the same flag of neutrality, aimed squarely at the standard that got there first. The last two times this argument played out, it set the terms for a generation. This one is being fought at agent speed.