The short answer
Five of enterprise software's biggest names — Google, Microsoft, Salesforce, Snowflake and ServiceNow — are lining up behind a shared agent-communication protocol as a counterweight to Anthropic and OpenAI, according to reporting by The Information in mid-July 2026. The battleground is the “plumbing” layer: the standards that decide how AI agents connect to enterprise data, tools and each other. Anthropic's Model Context Protocol (MCP) has quietly become the de-facto tool-connection standard over roughly 18 months, and the incumbents would rather not build the agent era on a rival's foundation.
The practical reading for engineering leaders: this is not one protocol beating another, it is a stack settling into layers — MCP-style tool access underneath, Agent2Agent (A2A) style agent-to-agent orchestration on top. If you are building AI agent systems, the winning move is not to pick a side but to design so you can adopt both and swap either without rewriting your business logic.
What actually happened?
In mid-July 2026, The Information reported under the headline “Google and Microsoft Team Up to Beat Back Anthropic and OpenAI” that a group of enterprise incumbents — Google, Microsoft, Salesforce, Snowflake and ServiceNow — are aligning behind a shared agent-communication protocol. Read the roster and the logic is obvious: Salesforce holds the customer data, Snowflake holds the analytical data, ServiceNow holds the workflows, and Google and Microsoft own the clouds all of it runs on. Between them they touch most of the enterprise data and processes that AI agents will need to act on.
What they are contesting is not a model or an app — it is the connective tissue. Over the last 18 months, Anthropic's Model Context Protocol (MCP) became the default way to plug an agent into tools and data, to the point that even Google adopted it across its services in December 2025. For companies whose strategic value is being the system of record, standardizing on a competitor's protocol at the layer that decides how agents reach enterprise data is uncomfortable. Backing an alternative at the agent coordination layer is how they answer it. Anyone building serious Anthropic-based or multi-vendor agent stacks now has to reason about two protocol layers, not one.
It is worth stating what is not confirmed. The reporting describes an alignment and shared intent, not a finished, single, co-branded specification shipping on a fixed date. Treat the specifics as directional, and treat the direction — enterprise incumbents rallying to own the agent-orchestration layer — as the real, verifiable story.
Why is this a layered fight, not a single winner?
The most common mistake in reading this story is treating it as “MCP vs A2A, one must lose.” They largely address different layers. MCP standardizes how a single agent connects to tools and data — files, databases, APIs, SaaS systems. Agent2Agent (A2A), the protocol the incumbents have coalesced around, standardizes how separate agents discover and talk to one another across platform and organizational boundaries, with signed agent cards and cryptographic verification of who is talking to whom. A realistic enterprise system uses both: MCP to reach the tools, A2A to coordinate the agents.
By adoption, MCP is not fading. Anthropic reports it has reached roughly 10,000 servers and about 97 million monthly SDK downloads, and A2A itself was already in production at around 150 organizations by April 2026. The fact that the largest incumbents felt they needed a rallying point at the coordination layer is a measure of how firmly MCP landed at the tool layer — not a sign it is being displaced. The honest description of where this is heading is a multi-protocol stack, with the fiercest competition over who governs the orchestration layer that sits above tool access. That layer matters because whoever defines how agents discover, coordinate and execute cross-functional work has real leverage over the enterprise stack.
Who governs the standards now?
Here is the twist that makes the “war” framing incomplete: A2A is now governed by the Linux Foundation's Agentic AI Foundation, and the same incumbents lining up commercially against Anthropic also sit inside that foundation alongside a broad roster of vendors. Everyone is cooperating on the shared standard in the foundation and competing hard in the market at the same time. For buyers, that combination is a feature, not a contradiction: a protocol under neutral, multi-vendor governance is far safer to build on than a proprietary interface a single vendor can change, meter or deprecate on its own timetable.
This is the part that should shape architectural decisions more than any vendor headline. When you evaluate an agent protocol, the governance model is a first-class criterion next to the technical fit. Foundation-governed standards lower your lock-in risk; single-vendor protocols raise it, however good the ergonomics feel on day one. The same discipline applies when you weave these agents into existing systems through GenAI integration work — the integration should sit behind an interface you own, so the protocol underneath stays replaceable.
What it means for US & EU software teams
Strip away the corporate maneuvering and three durable implications remain, none of which require you to pick a camp. The first is design for a layered, multi-protocol world. Treat tool access and agent-to-agent orchestration as two separate concerns, each behind its own adapter, so that adopting MCP for tools and A2A for coordination — or replacing either later — is a contained change rather than a re-platforming. Teams that hard-wire one vendor's protocol through their business logic will pay for it the first time the standard shifts, and in this space it will shift.
The second is weight governance as heavily as features. For US and EU teams especially, the protocols you build on carry compliance and continuity consequences: how agents authenticate to each other, how data crosses service and jurisdictional boundaries, and whether a critical dependency is controlled by one vendor or a neutral foundation. Prefer standards under foundation governance, and document the trust and data-flow model of every agent-to-agent link — regulators and security reviewers will ask, and cryptographically signed agent identity (as in A2A) is easier to defend than ad-hoc integrations.
The third is architectural and familiar from every fast-moving layer of the stack: own the part that is actually yours. The durable asset is not the wire protocol — it is your orchestration logic, your data governance, your evaluation and testing harness, and the domain knowledge encoded in how your agents behave. Keep those portable and vendor-neutral, and let the protocol be a component you can swap. Do that and a standards war between giants becomes something you observe and benefit from, rather than something that dictates a rebuild.
What to do now
Here is the shippable version. Treat the alliance news as a well-timed prompt to get your agent architecture portable, whatever protocols you end up using.
- Separate the layers. Put tool access (MCP-style) and agent-to-agent orchestration (A2A-style) behind distinct adapters, not tangled through your business logic.
- Adopt both where useful. These protocols are complementary; plan to use MCP for tools and A2A for coordination rather than forcing an either/or.
- Score governance, not just ergonomics. Prefer standards under a neutral foundation; flag any single-vendor protocol as a lock-in risk in your architecture review.
- Nail agent identity and data flow. Use signed, verifiable agent identity and document how data crosses service and jurisdictional boundaries for every agent link.
- Keep an exit hatch. Ensure you can change protocols without touching orchestration logic, evaluation harnesses or your source of truth for data.
- Pilot before you standardize. Prove the stack on one bounded, low-risk workflow before rolling a protocol choice across teams.
None of this is a verdict on any single vendor. Google, Microsoft, Salesforce, Snowflake, ServiceNow and Anthropic may all end up in your stack. The strategic signal is simply that the agent era's plumbing is being fought over right now — and the teams that stay portable through the fight are the ones who get to use the winners without being captured by them.
Frequently asked questions
What did Google, Microsoft and enterprise allies announce?
According to reporting by The Information in mid-July 2026, Google, Microsoft, Salesforce, Snowflake and ServiceNow are aligning behind a shared agent-communication protocol as a counterweight to Anthropic and OpenAI in enterprise AI. The vehicle they have coalesced around is Agent2Agent (A2A), a standard for agent-to-agent coordination across platforms and organizations, now governed by the Linux Foundation's Agentic AI Foundation. The move targets the “plumbing layer” that decides how AI agents connect to enterprise data, tools and each other — a layer where Anthropic's Model Context Protocol has become the de-facto standard over the past 18 months.
What is the difference between MCP and A2A?
They operate at different layers and are largely complementary rather than direct substitutes. Anthropic's Model Context Protocol (MCP) standardizes how a single agent connects to tools and data sources — files, databases, APIs and SaaS systems. Agent2Agent (A2A) standardizes how separate agents discover and communicate with one another across platforms and organizational boundaries, including signed agent cards with cryptographic verification. A realistic enterprise stack can use both: MCP for tool access and A2A for agent-to-agent orchestration. The competitive framing is really about which layer and which governance body sets the terms for the era.
Is MCP losing ground to the enterprise alliance?
Not by adoption numbers. Anthropic reports MCP has reached roughly 10,000 servers and about 97 million monthly SDK downloads, and Google adopted MCP across its services in December 2025. That the largest incumbents felt the need to rally around an alternative at the agent-to-agent layer is a signal of how firmly MCP landed at the tool layer, not evidence it is fading. The more accurate reading is a multi-protocol world settling into layers, with vendors competing to govern the orchestration layer above the tool layer.
Are these agent protocols open standards?
Increasingly, yes. A2A is now governed by the Linux Foundation's Agentic AI Foundation, and notably all five alliance members also participate in that foundation — cooperating on shared standards while competing commercially. MCP originated at Anthropic and has broad third-party adoption. Foundation governance reduces single-vendor control, which is exactly why it matters for buyers: standards under a neutral body are safer to build on than proprietary interfaces that one vendor can change or deprecate.
What should teams building AI agents do now?
Design for a layered, multi-protocol world instead of betting the architecture on one vendor. Treat tool access (MCP-style) and agent-to-agent orchestration (A2A-style) as separate, swappable layers behind your own interfaces. Prefer standards governed by a neutral foundation over proprietary ones, keep an adapter boundary so you can change protocols without rewriting business logic, and pilot on a bounded, low-risk workflow before standardizing. The durable asset is your own orchestration logic, data governance and evaluation harness — not the specific wire protocol underneath.
Sources
The Information — Google and Microsoft Team Up to Beat Back Anthropic and OpenAI
TNW — Google Cloud Next 2026: AI agents, A2A protocol, and the full-stack bet against OpenAI and Anthropic
Futurum Group — ServiceNow and Google Cloud's AI Agent Alliance
Linux Foundation — Agentic AI Foundation (A2A governance)