The "Protocol War" Anxiety—and a Verdict That Is Already In
When enterprises adopt AI agents, the question that stalls decision-makers most often is not technical but tribal: Anthropic's MCP, Google's A2A, and IBM's ACP appeared almost simultaneously, and betting on the wrong standard could mean rebuilding the entire integration investment. This "format war" anxiety is familiar from tech history—from videotape formats to browser engines, picking the losing side has always been costly. But by mid-2026, the shape of this contest is already quite clear, and the ending is gentler than most expected: the major protocols did not fight to the death; they settled into different layers, and some even merged.
Half of the conclusion first: MCP (Model Context Protocol) was released by Anthropic in November 2024 and formally donated on December 9, 2025 to the Agentic AI Foundation (AAIF), newly established under the Linux Foundation. The official figures at donation time were 97 million monthly SDK downloads and more than 10,000 active servers, and the AAIF platinum member list—AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft, OpenAI—captures nearly every competing giant at once. For anyone signing off on budgets, this means betting on MCP is no longer a gamble but catching up with a fait accompli. This article is not about how to wire up MCP; it is about the step before that: when multiple standards appear, what framework should an enterprise use to decide which to back, when to commit, and when to wait.
Understand the Layers: These Protocols Are Not Competing for the Same Seat
The "which one do we choose" anxiety mostly comes from assuming these protocols replace one another. In reality they belong to different layers of an agent system—more a stack than a substitution:
| Protocol | Layer It Solves | Governance Status | Key Figure |
|---|
| MCP | Agent ↔ external tools and data (vertical integration) | Linux Foundation AAIF (since Dec 2025) | Official registry listed roughly 9,600 servers as of May 2026 |
| A2A | Agent ↔ agent (cross-vendor horizontal collaboration) | Linux Foundation (since June 2025) | Over 150 organizations supporting it at its first anniversary in April 2026 |
| ACP | Agent communication (originated from IBM BeeAI) | Merged into A2A in August 2025 | No longer an independent option |
In other words, MCP lets a single agent connect to databases, APIs, and internal systems; A2A lets agents from different vendors discover each other's capabilities and delegate tasks. And ACP—the protocol most often invoked to manufacture "pick one of three" anxiety—announced its merger with A2A back in August 2025, with the Linux Foundation's BeeAI platform switching to A2A as well. The supposed three-way standoff is now, in practice, "one standard per layer": MCP for tools, A2A for collaboration.
Lay Out the Timeline and the Signals Are Easy to Read
Looking back at the adoption timeline, the pivotal turns were not any single launch event but the moments when competitors started adopting:
- March 2025: OpenAI announced MCP support in the Agents SDK and signaled that ChatGPT desktop and API support would follow; the Responses API went on to add remote MCP server support that May—a competitor adopting a protocol initiated by Anthropic was the first decisive signal.
- May 2025: Microsoft and GitHub announced at Build that they were joining the MCP steering committee and building MCP support directly into Windows 11.
- June 2025: Google donated A2A—announced only in April—together with its specification and SDKs to the Linux Foundation, voluntarily giving up single-vendor control.
- December 2025: Anthropic donated MCP to the AAIF, bringing it under neutral governance alongside Block's goose and OpenAI's AGENTS.md.
- April 2026: The Linux Foundation published A2A's first-year results: supporting organizations grew from fewer than 50 in April 2025 to more than 150, including AWS, Cisco, Google, IBM, Microsoft, Salesforce, SAP, and ServiceNow; Microsoft has integrated A2A into Azure AI Foundry and Copilot Studio, AWS added support via Amazon Bedrock AgentCore Runtime, and there are production deployments across multiple industries; v1.0 added Signed Agent Cards to close the cross-organization trust gap.
Adoption data has caught up too: Stacklok's 2026 software report found that 41% of surveyed software organizations were already running MCP servers in production (limited or broad), and cumulative MCP SDK downloads passed 150 million in 2026. None of these are a single vendor's marketing numbers—they are cross-camp, cross-verifiable signals.
Three Signals to Read Before Placing a Bet
Distilling that history into a reusable decision framework, whether a standard is worth backing comes down to three things:
- Breadth of adoption: when direct competitors all adopt the same protocol, no single vendor can hold it hostage. OpenAI, Microsoft, and Google successively backing MCP is exactly this signal.
- Neutral governance: handing a standard to a neutral foundation means no one company can unilaterally change the rules or revoke the license. MCP and A2A both now sit under the Linux Foundation—the most important guardrail for enterprises worried about lock-in.
- Reversibility: can the cost of the bet be recovered? Integrations built on open standards do not need rewriting when you switch models or vendors; bet on a closed platform's proprietary format and the downside risk is entirely yours.
The counter-signal matters just as much. A mature standard does not mean a mature ecosystem: a quantitative study of the MCP ecosystem analyzed more than 17,000 community-listed server records and found that only about half passed validity checks—meaning the other half were abandoned, placeholder, or otherwise unusable projects. Betting on a protocol and selecting a specific server are two different decisions—the former is now safe, the latter still needs case-by-case verification.
A Pragmatic Betting Strategy: Stake by Layer, Not All at Once
| Your Situation | The Bet to Place |
|---|
| Agents only need to connect to internal systems and APIs | Adopt MCP directly—this layer is settled, no need to wait |
| Multiple agents must collaborate across departments or vendors | Use MCP for tools while evaluating A2A for collaboration in parallel |
| Still evaluating, small scale | Back MCP first; watch A2A without deep commitment yet |
| Considering building a proprietary integration format | Stop—the maintenance and lock-in cost of a custom format now exceeds adopting the open standard |
The core principle: adopt decisively in the layer where the contest is settled (tool integration), stay compatible but defer heavy bets in the layer still converging (multi-agent collaboration), and only observe—without betting—in the layers that are still young (agent payments, agent interfaces). Stay sober about "backing the standard guarantees success" as well: Gartner predicted in June 2025 that over 40% of agentic AI projects will be canceled by the end of 2027, mainly due to escalating costs, unclear business value, and inadequate risk controls; its 2026 Hype Cycle places agentic AI at the Peak of Inflated Expectations, with only about 17% of organizations having actually deployed AI agents. Choosing the right protocol only means not betting on the losing side—project success still depends on use cases and governance. The real risk has never been "betting wrong on MCP"; it is waiting so long that your agents never connect to your own systems.
How Nerdtechnic Helps Enterprises Make This Call
In its AI adoption consulting service, Nerdtechnic helps enterprises break "which standard should we back" into decidable questions: inventory which systems your agents must reach, whether cross-agent collaboration is involved, which integrations are worth doing now, and which should wait for standards to settle. We advocate adopting open protocols governed by neutral foundations—like MCP—at the tool-integration layer, so your investment is not held hostage by any single vendor and integrations need not be rewritten every time you switch models. The point of betting on standards is not guessing who wins, but placing your stake in the layer you are least likely to regret.
References
- Anthropic, "Donating the Model Context Protocol and establishing the Agentic AI Foundation," 2025. Source
- Anthropic, "Introducing the Model Context Protocol," 2024. Source
- Linux Foundation, "Linux Foundation Announces the Formation of the Agentic AI Foundation," 2025. Source
- Ginger Labs, "Best MCP Servers in 2026" (citing roughly 9,652 servers on the official registry as of May 2026). Source
- Linux Foundation, "A2A Protocol Surpasses 150 Organizations, Lands in Major Cloud Platforms, and Sees Enterprise Production Use in First Year," 2026. Source
- Google, "Google Cloud donates A2A to Linux Foundation," 2025. Source
- LF AI & Data, "ACP Joins Forces with A2A under the Linux Foundation's LF AI & Data," 2025. Source
- OpenAI Developer Community, "Introducing support for remote MCP servers, image generation, Code Interpreter, and more in the Responses API," 2025. Source
- Microsoft, "Microsoft Build 2025: The age of AI agents and building the open agentic web," 2025. Source
- Stacklok, "State of Model Context Protocol in Software 2026," 2026. Source
- Infosecurity Magazine, "Systemic Flaw in MCP Protocol Could Expose 150 Million Downloads," 2026. Source
- Guo et al., "A Measurement Study of Model Context Protocol Ecosystem" (arXiv:2509.25292), 2025. Source
- MarTech (citing Gartner's June 2025 press release), "Gartner: 40% of agentic AI projects will fail, making humans indispensable," 2026. Source
- xpander.ai (citing Gartner's 2026 Hype Cycle for Agentic AI), "Gartner Hype Cycle for Agentic AI: What It Means for AI Agent Development Platforms," 2026. Source
Industry Trends
