Agentic Web Standards: MCP, A2A, NLWeb & AGENTS.md

Just as the early web required shared protocols like HTTP and HTML to unify disparate systems, the emerging agentic web now demands its own foundational standards. Without them, the ecosystem of AI agents risks the same fragmentation that nearly splintered the World Wide Web in its infancy. This article examines the four core protocols establishing that critical infrastructure layer, explaining their function, backing, and practical implications for businesses.

The necessity for standards is clear. In the 1990s, the World Wide Web Consortium (W3C) brought order by defining universal rules for data transport and structure. Today, AI faces an identical challenge. Every major technology firm is developing agents that must connect to tools, data, and each other. Without common protocols, businesses would face an unsustainable M x N problem, forced to build separate integrations for every combination of AI model and external system.

The collaborative response to this challenge is unprecedented. In December 2025, the Linux Foundation announced the Agentic AI Foundation (AAIF), a vendor-neutral body for governance. Its platinum members include direct competitors like AWS, Anthropic, Google, Microsoft, and OpenAI. These rivals are cooperating on shared infrastructure because they recognize that proprietary standards would stifle the entire ecosystem’s growth.

First among these new standards is the Model Context Protocol (MCP). Think of MCP as a universal adapter, a USB-C port for AI applications. It provides a standardized interface for connecting AI systems to external data sources, tools, and workflows. Before MCP, each platform required its own custom integration. Now, building one MCP server makes resources accessible to all compatible AI systems. Launched by Anthropic in late 2024, adoption exploded, reaching 97 million monthly SDK downloads within a year. Support from OpenAI, Google, and Microsoft transformed it from an experiment to a de facto industry standard in just 12 months. For businesses, this means a single connection can allow any major AI assistant to query real-time inventory, CRM data, or pricing information.

While MCP connects agents to tools, the Agent2Agent protocol (A2A) enables agents to communicate with each other. As enterprises deploy agents from different vendors for CRM, IT, and billing, these systems need to discover capabilities and collaborate. A2A solves this through Agent Cards, JSON documents that act as digital business cards, published at a standard web address. Launched by Google in April 2025 and swiftly donated to the Linux Foundation, A2A gained support from over 150 organizations, including Salesforce, SAP, and ServiceNow, within months. This rapid enterprise adoption signals a pressing need for multi-vendor agent coordination.

For website owners, NLWeb (Natural Language Web) is arguably the most immediately relevant protocol. Developed by Microsoft and spearheaded by R. V. Guha, the creator of Schema.org, NLWeb turns any site into a natural language interface. It leverages existing structured data like Schema.org markup to let users and AI agents query content conversationally. Crucially, every NLWeb instance functions as an MCP server, making a website’s content instantly accessible to the entire MCP ecosystem. Early adopters like Eventbrite, Shopify, and Tripadvisor demonstrate its value for content-rich sites, allowing queries like “find family-friendly restaurants in Barcelona with outdoor seating” to return structured, actionable data.

The final piece is AGENTS.md, a standardized Markdown file that guides AI coding agents. It provides project-specific instructions on architecture, conventions, and testing, ensuring consistent, high-quality output. Adopted by over 60,000 open-source projects since its mid-2025 release, it is supported by tools like GitHub Copilot and Claude Code. For development teams, implementing AGENTS.md is a straightforward step toward better quality control, reducing bugs and ensuring AI-generated code aligns with internal standards.

These four protocols form a complementary stack, not a competitive set. MCP provides the plumbing for tool access. A2A enables agent communication. NLWeb makes website content agent-ready. AGENTS.md guides development. The AAIF governs them all. The parallel to the original web is striking, but the speed is unprecedented. Adoption is occurring at internet velocity, driven by the immediate needs of the companies building the agents themselves.

Businesses should assess their readiness strategically. If you have invested in Schema.org markup, you are already positioned for NLWeb. If you maintain APIs or internal tools, consider their MCP accessibility to serve any AI platform. Organizations using multi-vendor agent workflows should monitor A2A for seamless coordination. Development teams using AI coding tools can benefit immediately from adopting AGENTS.md.

The overarching message is that the agentic web is being built on open standards. Understanding these protocols now is crucial as AI agents become a primary interface for user interaction. The infrastructure is being laid at remarkable speed, mirroring the collaborative moment that unified the early web and enabled its explosive growth.