The Protocol Wars Deciding the Agentic Web

Published: September 23, 2026 | Category: Technology | By Mahesh

Connecting an AI agent to a tool is now a solved problem. Connecting an AI agent built by one company, on one model, to a second agent built by a different company on a different model, was not solved until 2025, and the two protocols that emerged to solve it are on track to become as invisible and as foundational as TCP/IP or HTTP, protocols most people have never heard of yet depend on every day.^[1] Anthropic released the Model Context Protocol, MCP, in November 2024 to standardize how an individual AI agent reaches external tools and data. Google introduced a separate protocol, Agent2Agent, or A2A, in April 2025, specifically to let independent agents built by different vendors talk to each other as peers, and donated it to the Linux Foundation within months.^[1] By early 2026, MCP was seeing more than 110 million downloads a month, and A2A had crossed 150 supporting organizations including Amazon Web Services, Microsoft, Cisco, Salesforce, SAP, and ServiceNow, reaching what the Linux Foundation itself calls production-grade enterprise adoption.^[2]

This is the founding pillar for a new Depth Grid series looking at the infrastructure layer underneath the agentic AI boom, the plumbing question almost no mainstream coverage of AI actually asks: once every company has an AI agent, what set of rules decides how those agents find, trust, and cooperate with each other, and which company's rules become the ones the rest of the industry has to build around. Depth Grid's earlier reporting this year traced why the AI moat is moving from the model to the workflow and documented why enterprise AI agents are where the big venture checks are going. This piece goes one layer deeper, into the specific technical standards battle that will determine who actually controls the rails those agents run on for the next five to ten years, long after this particular funding cycle's headlines have faded.

The Problem Neither Protocol Alone Can Solve

Before MCP and A2A existed, every AI integration was custom, bespoke code. Connecting an agent to a database required writing custom tool definitions. Connecting that same agent to a second, different agent required an entirely separate layer of custom code. Surfacing the result in a user interface required yet more custom work. None of it scaled, and every new integration meant starting from scratch.^[3] MCP solved the first half of that problem specifically: it standardizes how an AI model connects to external tools, data sources, and services, using a client-server architecture in which an MCP server exposes resources and tools, and an MCP client, the agent itself, discovers and uses whatever that server exposes.^[3] The protocol is frequently described, including by its own community, as "USB-C for AI," a single, universal connector that replaces a drawer full of incompatible cables.^[4]

MCP was never designed to solve the second half of the problem, however, and its own design philosophy is explicit about that boundary: MCP standardizes plumbing and deliberately stays out of opinions about agent architecture or how independent agents should negotiate with one another.^[1] That gap is precisely what A2A was built to fill. Industry analysis of the current stack is direct about the distinction: MCP gives an agent its tools, A2A gives multiple agents a shared language to talk to one another, and a serious production deployment needs both, since an enterprise's Salesforce agent, its ServiceNow agent, and a custom agent built in-house will not all come from one vendor, run one model, or live in one cloud.^[1] Picking only one protocol and ignoring the other leaves half the stack mute: agents that can act but cannot collaborate, or agents that can negotiate with each other but have no hands to actually do anything.^[1]

MCP: the USB-C Layer, and How Fast It Actually Grew

MCP's adoption curve is one of the fastest of any developer protocol in recent memory, and the specific timeline is worth tracing because the speed itself is part of the story. Anthropic introduced MCP in late 2024 to largely polite, contained interest, treated at first as one vendor's own integration scheme among many similar efforts.^[5] That changed within months: by early 2025, MCP adoption crossed what industry analysts describe as a genuine tipping point, as other major model providers and development frameworks embraced it, and the ecosystem realized a tool built once could now serve every agent built on top of it, not just Anthropic's own.^[5] The protocol's own maintainers reported over 110 million monthly downloads by early 2026, and its latest formal specification, dated November 25, 2025, has evolved considerably beyond its original scope, now supporting distributed execution, stateless transport addressing early load-balancing limitations, and a formal security framework.^[2]

MCP's inspiration and technical lineage matter for understanding why it scaled so quickly: it draws directly on the Language Server Protocol, an existing, well-understood standard originally built to let code editors support many programming languages without each editor needing custom logic for every language.^[6] Applying that same "write once, serve everyone" logic to AI tool access, rather than inventing an entirely novel architecture from scratch, gave developers a genuinely familiar mental model to adopt quickly. The protocol's technical implementation, using JSON-RPC 2.0 over a defined transport layer with an initialization handshake and capability negotiation, has since achieved integration with major LLM providers well beyond Anthropic itself, including OpenAI and Google, exactly the kind of cross-vendor adoption that separates a genuine industry standard from a single company's proprietary tool.^[6]

A2A: From Google Project to Linux Foundation Standard

Google's own framing of A2A from the outset positioned it deliberately as MCP's complement rather than its competitor: tools versus peers, a layered-stack framing that entered the industry discourse from the protocol's earliest days rather than emerging later as a diplomatic reframing after competitive tension surfaced.^[5] Google introduced A2A in April 2025 with dozens of partners already signed on, and moved quickly into what industry analysis calls the institutional phase: Google donated A2A to the Linux Foundation within months of its launch, and a rival agent-communication effort originally developed independently by IBM Research, called the Agent Communication Protocol, subsequently merged into A2A rather than continuing as a separate, competing standard.^[5]

Sources: Linux Foundation via Beam.ai and DEV Community protocol-stack analysis, 2026.^[1][2][5]

By April 2026, A2A reached version 1.0, backed by more than 150 organizations and formally integrated into Amazon Web Services, Microsoft, and Google's own cloud platforms, positioning it as the de facto standard for inter-agent communication specifically within enterprise contexts.^[2] The Linux Foundation's own founding partner list for the project reads like a cross-section of the entire enterprise software industry: Amazon Web Services, Cisco, Google, Microsoft, Salesforce, SAP, and ServiceNow.^[2] That breadth of institutional backing, spanning direct cloud competitors who would normally have every commercial incentive to build rival, proprietary standards instead, is itself a meaningful signal about how seriously the industry has decided this particular coordination problem needs a single shared answer rather than a fragmented one.

The Layered Stack: Six Protocols, Six Different Jobs

MCP and A2A are the two most widely adopted protocols, but they are not the only ones, and understanding where each one fits within a broader layered stack clarifies why this is a genuinely complex infrastructure question rather than a simple two-horse race. Current industry analysis identifies a full stack forming around distinct layers, each answering a different question: tools, then the web, then agent-to-agent communication, then semantics, then payments, then the human interface, with identity and security woven through every layer as cross-cutting concerns.^[7] Beyond MCP and A2A, additional protocols occupy specific, narrower niches within this stack: AG-UI and A2UI address how an agent's output actually surfaces in a human-facing interface, AP2 addresses agent-initiated payments, and X42 addresses a further specialized layer, while IBM's own Agent Communication Protocol, before its merger into A2A, and the separately developed Agent Network Protocol both proposed alternative approaches to peer-to-peer agent discovery using decentralized identity standards.^[7]

The Agent Network Protocol specifically illustrates a genuinely different philosophical approach still competing for adoption within this same broad space: rather than building on the existing internet's human-oriented infrastructure, ANP proposes a layered architecture incorporating World Wide Web Consortium decentralized identity standards and semantic web principles, built on the premise that the modern internet, optimized for human interaction, is structurally suboptimal for the low-latency, API-native communication autonomous agents actually require.^[8] Whether that more radical, internet-native-for-agents approach gains meaningful adoption against the more incremental, build-on-existing-infrastructure approach MCP and A2A both represent is itself one of the genuinely open questions this protocol stack has not yet resolved, and is a dynamic worth returning to as this series continues.

Why This Is a Real Standards War, Not Settled Cooperation

Public messaging from both Anthropic and Google consistently frames MCP and A2A as complementary rather than competing, and at a pure technical level, that framing is accurate: the two protocols genuinely operate at different layers and are frequently deployed together in the same production system. But treating the underlying dynamic as fully settled cooperation, with no real competitive tension, misses an important nuance industry practitioners themselves flag directly: most enterprises still treat the two as competing standards in practice, even when the protocols' own designers insist they are not, because the conflation is genuinely understandable given how the two arrived within the same eighteen-month window, both carry the word "protocol" in their name, and both are frequently marketed using the same promise of interoperability for AI agents.^[1]

The more consequential competitive question is not MCP versus A2A specifically, both of which now appear likely to coexist as complementary layers, but which company's broader agent ecosystem, and which company's specific implementation choices within each protocol layer, becomes the default enterprises build around. A protocol being open and vendor-neutral in its formal governance, as both MCP and A2A now are, does not fully neutralize the first-mover advantage the company that originated it retains, in developer mindshare, in the depth of its own internal tooling built around the standard, and in its ability to shape the protocol's future evolution through continued influence within the governing body, even after formal control has been handed to a neutral foundation. This is precisely the dynamic that makes this a genuine standards war worth tracking closely, even though neither Anthropic nor Google frames it that way publicly.

What History Says About Who Wins a Protocol War

Standards battles are not new to the technology industry, and the historical pattern across prior infrastructure wars offers a genuinely useful, durable lens for thinking about how this one is likely to resolve, regardless of which specific protocols happen to be involved this time. The internet's own foundational protocols, TCP/IP, HTTP, and later HTML and CSS, all won not because they were technically superior to every rival at the moment of adoption, but because they achieved broad enough, early enough cross-vendor buy-in that building on top of them became more economically rational for every subsequent participant than building a competing, isolated alternative. Once a protocol crosses that threshold, network effects compound quickly: more developers build tools for the dominant standard, which makes it more useful, which attracts more developers, in a self-reinforcing cycle that becomes progressively harder for a technically comparable but less-adopted rival to break.

MCP and A2A both appear to be crossing exactly that threshold already, based on the adoption figures cited throughout this piece, which is precisely why this protocol layer deserves the same kind of durable, infrastructure-level attention Depth Grid's earlier reporting gave to the physical compute and power constraints underlying AI's growth. A grid interconnection queue or a chip packaging shortage is a visible, physical bottleneck anyone can photograph. A protocol standard winning quiet, cumulative adoption among the seven or eight companies that together control most of enterprise cloud infrastructure is a far less visible event, with no ribbon-cutting ceremony or press conference, but it is arguably just as consequential for who captures the economic value of the agentic AI buildout over the coming decade, since the company or coalition that controls the default rails everyone else builds on top of typically captures a disproportionate, durable share of the value flowing through that infrastructure, largely independent of which individual companies win or lose in the applications built on top of it.

What This Means for the Next Five Years

Treat protocol choice as a genuine architectural decision, not a minor implementation detail. Given that Gartner projects 40 percent of enterprise applications will feature task-specific AI agents by the end of 2026, up from under 5 percent in 2025, an eightfold jump in a single year, any organization building agentic systems now should treat its choice of which protocols to build on as a decision with multi-year consequences, evaluated with the same seriousness historically reserved for choosing a core database or cloud provider, not a detail left to whichever engineer happened to read about a given protocol first.

Expect to run both MCP and A2A together, not to have to choose one exclusively. Given that the two protocols operate at genuinely different layers of the stack, tool access versus agent-to-agent coordination, a production-grade agentic deployment connecting to external systems built by multiple vendors should plan to implement both rather than treating the choice as either-or, since skipping either layer leaves a meaningful category of real-world agent coordination unsupported.

Watch the governance structure, not just the download numbers, when assessing a protocol's long-term durability. A2A's donation to the Linux Foundation and IBM's decision to merge its own competing effort into A2A rather than continuing a rival standard are both stronger long-term durability signals than raw adoption numbers alone, since a protocol governed by a genuinely neutral, multi-stakeholder foundation is structurally less vulnerable to a single company later changing terms unilaterally than one still controlled entirely by its originating vendor.

Common Questions

This analysis is editorial commentary based on publicly available sources cited above. It is not technical, investment, or engineering advice. Protocol adoption figures, version numbers, and organizational backing cited reflect data available as of publication and change frequently as this standards landscape evolves; verify current figures directly with the Linux Foundation, Anthropic, and Google before making architecture decisions based on this information.

Sources

1. Beam.ai, "Agent2Agent vs MCP: 2 Protocols Your 2026 Stack Needs," citing Linux Foundation adoption data and Gartner August 2025 forecast, June 23, 2026. Link
2. Agentverse Platform, "Infrastructure for the Agentic Web: Gap Analysis and Architecture," citing MCP Core Maintainers 2026 data and Linux Foundation A2A v1.0 details. Link
3. MindStudio, "Six Agent Protocols Every AI Builder Needs to Know in 2026," May 20, 2026. Link
4. AgentRob research paper, "From Virtual Forum Agents to Hijacked Physical Robots," citing MCP technical architecture. Link
5. DEV Community, "The State of Agentic AI Standards in 2026: MCP, A2A, WebMCP, OSI, and the Protocol Stack Taking Shape," July 7, 2026. Link
6. From LLM Reasoning to Autonomous AI Agents, "A Comprehensive Review," citing MCP's Language Server Protocol lineage. Link
7. DEV Community, "The State of Agentic AI Standards in 2026," July 7, 2026 (layered stack detail). Link
8. A Survey of Agent Interoperability Protocols, "Model Context Protocol (MCP), Agent Communication Protocol (ACP), Agent-to-Agent Protocol (A2A), and Agent Network Protocol (ANP)," arXiv. Link

Read More on Depth Grid

• Why the AI Moat Is Moving From the Model to the Workflow
• Enterprise AI Agents Are Where the Big Checks Are Going
• The Compute-and-Power Bottleneck: Why AI's Next Decade Will Be Decided by Electrons, Not Algorithms
• The Great AI Layoff Reversal: Why Gartner Says a Third of AI-Cut Jobs Will Come Back by 2029

Article by Mahesh | Depth Grid