In the spring of 2025, Anthropic published the specification for the Model Context Protocol — a technical standard for connecting AI agents to external tools, databases, and APIs. Within twelve months, the protocol had accumulated over 18,000 community-indexed server implementations and tens of millions of monthly SDK downloads. By mid-2026, it had been donated to the Linux Foundation's newly formed Agentic AI Foundation and adopted as the de facto standard for agent-to-tool communication across the industry.
The speed of this adoption is remarkable. But what is more remarkable — and more consequential — is what it reveals about the structural dynamics of the emerging agentic AI ecosystem. For the first time in the history of AI development, the industry is converging on a shared protocol stack: a layered architecture of open standards that governs how agents discover capabilities, communicate with tools, coordinate with one another, and establish trust across organisational boundaries.
This convergence is not accidental. It reflects a recognition, hard-won through years of fragmented proprietary integration, that the value of agentic AI systems is fundamentally network-dependent. An agent that can only communicate with tools built by its own vendor is a closed system. An agent that can communicate with any tool, any service, and any other agent — regardless of vendor, jurisdiction, or technical stack — is infrastructure. The difference between these two architectures is the difference between a proprietary telephone network and the internet.
This analysis examines the emerging protocol stack for agentic AI interoperability, the governance structures that are beginning to form around it, and the implications for digital sovereignty, competitive dynamics, and the long-term architecture of the intelligent economy.
The Three-Layer Protocol Stack
By 2026, the agentic AI protocol landscape has coalesced around a three-layer architecture, each layer addressing a distinct dimension of the interoperability challenge.
Layer One: Tool Integration — The Model Context Protocol
The Model Context Protocol (MCP) addresses the "agent-to-tool" problem: how does an AI agent access external capabilities — databases, file systems, APIs, web services — in a standardised, secure, and composable way?
MCP's architecture is built on JSON-RPC 2.0, supporting both local (stdio) and production (Streamable HTTP) transport modes. It defines three primitive types: Resources (data that agents can read), Tools (functions that agents can call), and Prompts (templates that structure agent interactions). The 2025-2026 protocol updates introduced bidirectional capabilities — allowing servers to request LLM completions and elicit structured user input — transforming MCP from a simple tool-calling interface into a full-featured context management system.
The protocol's adoption trajectory is instructive. Within a year of publication, MCP had achieved the kind of network effects that typically take decades to accumulate in enterprise software standards. This speed reflects both the quality of the protocol design and the timing of its release: it arrived precisely when the industry was experiencing the pain of fragmented, proprietary integration at scale.
Layer Two: Agent Coordination — The Agent-to-Agent Protocol
The Agent-to-Agent (A2A) protocol, originally developed by Google and subsequently donated to the Linux Foundation, addresses the "agent-to-agent" problem: how do autonomous agents discover one another's capabilities, delegate tasks, and coordinate complex multi-step workflows across organisational boundaries?
A2A's architecture centres on "Agent Cards" — machine-readable manifests published at a standardised endpoint (/.well-known/agent.json) that describe an agent's capabilities, authentication requirements, and supported task types. Unlike the stateless nature of tool-calling in MCP, A2A manages the full lifecycle of collaborative work, tracking task states (submitted, working, completed, failed) and supporting multi-turn dialogue and streaming progress updates via Server-Sent Events.
The convergence on a shared protocol stack for agentic AI represents the most significant standardisation event in the history of AI infrastructure — one that will shape the competitive dynamics of the intelligent economy for decades.
The convergence on a shared protocol stack for agentic AI represents the most significant standardisation event in the history of AI infrastructure — one that will shape the competitive dynamics of the intelligent economy for decades.
The governance significance of A2A extends beyond its technical architecture. By establishing a standard mechanism for cross-organisational agent discovery and delegation, it creates the technical foundation for what researchers are beginning to call the "agentic web" — an ecosystem in which autonomous agents from different vendors, operating under different governance frameworks, can collaborate on complex tasks without requiring bespoke integration work.
Layer Three: REST-Native and Interface Protocols
The third layer of the protocol stack addresses specific transport and interface requirements that MCP and A2A do not fully cover. The Agent Communication Protocol (ACP), originally developed by IBM's BeeAI initiative, provides a pure REST-native approach to agent messaging — using standard HTTP verbs and OpenAPI schemas to enable integration with legacy HTTP toolchains without requiring specialised SDKs.
WebMCP, developed collaboratively by Google and Microsoft through the W3C, extends the MCP model to the browser environment. It allows websites to expose functionality as callable tools through a consensual, structured interface — effectively replacing brittle web-scraping techniques with a standardised, permission-based interaction model. The W3C AI Agent Protocol Community Group, which is driving the development of open web standards for agent interoperability, reached a significant milestone in June 2026 with the adoption of its AI Agent Memory Interoperability Community Group v1.0 charter.
The Semantic and Identity Layers
Protocol convergence at the transport and coordination layers is necessary but insufficient for enterprise-grade interoperability. Two additional layers — semantic interoperability and identity — are emerging as critical complements to the core protocol stack.
Semantic Interoperability: The Open Semantic Interchange
The Open Semantic Interchange (OSI) specification addresses a problem that becomes acute as agents move into enterprise data environments: the "confident wrongness" of AI systems that can retrieve data but cannot reliably interpret it. When an agent queries a business intelligence system, the raw data it receives may be technically correct but semantically ambiguous — the same metric defined differently across different systems, the same dimension labelled inconsistently across different data sources.
OSI provides a vendor-neutral specification for business metrics, dimensions, and relationships, enabling agents to interpret data accurately across different BI tools and data catalogues. It represents a recognition that protocol-level interoperability is not sufficient without semantic interoperability — the shared understanding of what data means, not merely how it is formatted and transmitted.
Identity and Trust: OAuth 2.1 and Decentralised Identifiers
The identity layer of the agentic protocol stack is perhaps the most consequential for questions of sovereignty and governance. How does an agent prove its identity to another agent? How does a tool verify that the agent requesting access has the authority to do so? How are credentials managed across organisational boundaries without creating centralised points of failure or control?
The emerging consensus combines two complementary approaches. OAuth 2.1 with Resource Indicators (RFC 8707) provides the short-term authentication layer, preventing credential leakage in multi-agent workflows by ensuring that tokens issued for one resource cannot be used to access another. W3C Decentralised Identifiers (DIDs) provide the long-term trust layer, enabling agents to establish verifiable identities that are not dependent on any single centralised authority.
The IETF is also contributing to this layer, with active work on protocol-level interoperability across independently administered internet domains — addressing the discovery, authorisation, delegation, and trust establishment requirements that arise when agents operate across jurisdictional and organisational boundaries.
Governance and the Linux Foundation Model
The governance of the agentic protocol stack — who controls the standards, who participates in their development, and who can veto changes — will determine whether the intelligent economy is built on open infrastructure or proprietary lock-in.
The governance structure that has emerged around the agentic protocol stack is as significant as the technical architecture itself. The Linux Foundation's Agentic AI Foundation (AAIF) now serves as the steward for both MCP and A2A — a governance model that has important implications for the long-term independence and neutrality of the protocol stack.
The Linux Foundation model is not without precedent. The Foundation has successfully stewarded critical open-source infrastructure — the Linux kernel, Kubernetes, the OpenAPI Specification — through periods of intense commercial competition, maintaining the neutrality and openness of these standards while enabling commercial adoption at scale. The application of this model to agentic AI protocols represents a deliberate choice to treat the protocol stack as shared infrastructure rather than proprietary competitive advantage.
The governance of the agentic protocol stack — who controls the standards, who participates in their development, and who can veto changes — will determine whether the intelligent economy is built on open infrastructure or proprietary lock-in.
This choice is not universally shared. Several major technology companies have maintained proprietary agent frameworks alongside their contributions to open standards — a strategy that preserves optionality while participating in the standardisation process. The tension between open standards and proprietary differentiation is a recurring dynamic in technology infrastructure, and the agentic AI ecosystem is unlikely to resolve it definitively.
What the Linux Foundation model does provide is a governance structure that is, at least in principle, resistant to capture by any single commercial interest. The Foundation's membership model, which includes both large corporations and smaller organisations, creates a degree of checks and balances that purely corporate governance structures lack. Whether this model will prove sufficient to maintain the neutrality of the protocol stack as commercial stakes increase remains to be seen.
Sovereignty Implications
The emergence of a shared agentic protocol stack has profound implications for digital sovereignty — at both the national and organisational level.
National Sovereignty and Protocol Governance
For nations pursuing sovereign AI strategies, the protocol layer represents a critical but often overlooked dimension of the sovereignty question. A nation that deploys AI systems built on open, internationally governed protocols retains the ability to switch vendors, audit implementations, and participate in the governance of the standards that shape its digital infrastructure. A nation that deploys AI systems built on proprietary protocols is dependent on the continued goodwill and commercial viability of the vendor that controls those protocols.
The CNAS Sovereign AI Index documents over 185 sovereign AI initiatives globally, with the majority focused on infrastructure — data centres, compute, and cloud sovereignty. But infrastructure sovereignty without protocol sovereignty is incomplete. A sovereign data centre running proprietary agent frameworks is not meaningfully more sovereign than one running proprietary cloud services.
The European Union's approach to this challenge is instructive. The EU's digital sovereignty strategy explicitly prioritises open standards and interoperability as mechanisms for reducing dependence on non-European technology providers. The EURO-3C initiative, which integrates telco, edge, and cloud resources through shared standards, represents an attempt to build federated infrastructure that is sovereign by design — not because it excludes foreign technology, but because it is built on standards that no single foreign entity controls.
Organisational Sovereignty and the Interoperability Imperative
At the organisational level, the protocol stack question is increasingly framed in terms of "managed interdependence" — a concept developed by the Brookings Institution and others to describe the realistic middle ground between full-stack sovereignty and complete dependence on external providers.
Organisations that build their agentic AI systems on open protocols retain the ability to switch components, audit implementations, and integrate with the broader ecosystem of tools and services. Organisations that build on proprietary frameworks accumulate technical debt in the form of switching costs — costs that grow as the systems become more deeply integrated into operational workflows.
For organisations and nations pursuing digital sovereignty, the protocol layer is not a technical detail — it is a strategic imperative.
The Linux Foundation's survey data suggests that 90% of organisations view open-source AI as essential to achieving meaningful sovereignty over their AI systems. This finding reflects a growing recognition that the protocol layer — not just the model layer — is a critical dimension of the sovereignty question.
The Competitive Dynamics of Protocol Convergence
The convergence on a shared protocol stack does not eliminate competition — it shifts its locus. When the transport and coordination layers are standardised, competition moves to the layers above and below: the quality of models, the richness of tool ecosystems, the sophistication of orchestration frameworks, and the depth of domain-specific capabilities.
This dynamic is familiar from the history of internet infrastructure. The standardisation of TCP/IP did not eliminate competition in networking — it created the conditions for competition at higher layers of the stack, ultimately producing the extraordinary diversity of applications and services that characterise the modern internet. The standardisation of the agentic protocol stack is likely to produce a similar dynamic: a shared foundation that enables competition at the application layer, rather than at the infrastructure layer.
For organisations building on this stack, the strategic implication is clear: competitive advantage will increasingly derive from the quality of the agents and tools built on top of the protocol stack, not from proprietary control of the stack itself. This is a fundamentally different competitive logic from the one that has characterised the AI industry to date — one that rewards capability and domain expertise over infrastructure control.
Outstanding Challenges
The emergence of a shared agentic protocol stack is a significant development, but it does not resolve all of the interoperability challenges facing the ecosystem. Several critical issues remain.
Security and trust remain incompletely addressed. The OAuth 2.1 and DID-based identity layer provides a foundation, but the security properties of multi-agent systems — where a compromised agent can potentially propagate malicious actions across an entire workflow — are not yet fully understood or mitigated. The IETF's work on cross-domain agent interoperability is addressing some of these challenges, but the field is still in early stages.
Semantic interoperability remains a significant gap. The OSI specification addresses the enterprise data interpretation problem, but the broader challenge of ensuring that agents from different vendors, trained on different data, with different internal representations, can reliably communicate about complex real-world concepts is far from solved.
Governance at scale is an open question. The Linux Foundation model has worked well for previous generations of open-source infrastructure, but the agentic AI ecosystem is developing at a speed and scale that may strain existing governance mechanisms. As the commercial stakes increase, the pressure on the neutrality of the protocol stack will intensify.
Conclusion
The convergence on a shared protocol stack for agentic AI — MCP for tool integration, A2A for agent coordination, ACP for REST-native messaging, and complementary semantic and identity layers — represents the most significant standardisation event in the history of AI infrastructure. It creates the technical foundation for an agentic web in which autonomous agents can discover, communicate, and collaborate across organisational and jurisdictional boundaries.
The governance of this stack — who controls the standards, who participates in their development, and who can veto changes — will determine whether the intelligent economy is built on open infrastructure or proprietary lock-in. The Linux Foundation model provides a governance structure that is, at least in principle, resistant to capture by any single commercial interest. Whether it will prove sufficient to maintain the neutrality of the protocol stack as commercial stakes increase is the central governance question of the agentic era.
For organisations and nations pursuing digital sovereignty, the protocol layer is not a technical detail — it is a strategic imperative. The ability to participate in, audit, and if necessary fork the standards that govern agentic AI infrastructure is a precondition for meaningful sovereignty in the intelligent economy.
