{"slug": "mcp-and-a2a-building-the-agentic-internet", "title": "MCP and A2A: building the agentic internet", "summary": "Anthropic's Model Context Protocol (MCP) and Google's Agent-to-Agent Protocol (A2A) are standardizing how AI agents discover tools, call services, and coordinate across systems, forming the infrastructure for an agentic internet where agents act on behalf of humans. MCP connects agents to tools and data via a client-server model, while A2A enables agents to delegate tasks to other agents using standard web infrastructure like HTTP and JSON-RPC 2.0. Together, these protocols address the need for interoperability in multi-agent systems.", "body_md": "Most of the conversation about AI agents focuses on what they can do: browse, research, transact, automate. Less attention has gone to the *how*. How do agents discover tools, call services, and coordinate with each other across systems they didn't build and vendors they don't control?\n\nThat *how* is now standardizing, and it matters to anyone building infrastructure that agents will use. The two protocols doing most of the work are the Model Context Protocol (MCP) from Anthropic and the Agent-to-Agent (A2A) Protocol from Google.\n\nThis overview will help you understand what each one does and how they work together to build the infrastructure for the agentic internet.\n\n## What is the agentic internet?\n\nIf the internet is infrastructure for human action, then the agentic internet is what happens when the main actors on that infra aren't people but agents acting on their behalf - doing the searching, authenticating, and transacting without humans getting involved at every step.\n\nThis is distinct from the agentic web, which is just the browser-facing layer: websites structured so agents can read and act on them. The agentic internet is the full stack underneath: the protocols that govern how agents find each other and communicate.\n\n## MCP - connecting agents to tools and data\n\nOne of those protocols is [MCP](https://modelcontextprotocol.io/docs/getting-started/intro), which was released in November 2024 because every AI-tool integration was bespoke: different APIs, auth, and data formats. Developers building agents that needed to call external services were writing custom connectors for each one. MCP standardized that.\n\nThe model is client-server. The agent is the client. The tool, data source, or service is the server. The agent calls a specific capability and gets a result back.\n\n*To connect your AI agent with tens of thousands of web tools, use Apify MCP server*\n\n## A2A - letting agents delegate to other agents\n\nMCP solves how an agent accesses a capability. The [Agent-to-Agent Protocol](https://a2aprotocol.ai/docs/) was launched in April 2025 to address a different problem. As enterprises deploy more agents, they increasingly operate in silos: separate systems and vendors, and no common way to coordinate. A2A is how agents work together across those boundaries.\n\nIn MCP, the calling agent decides exactly which tool to invoke and manages the entire workflow. In A2A, an orchestrator agent hands a task to a sub-agent and trusts it to figure out how to get the job done. The sub-agent has its own reasoning, tools, and workflow. The orchestrator doesn't manage the steps, but delegates the outcome. Agents can also coordinate as peers, sharing goals and dividing work without a fixed hierarchy.\n\nThe technical foundation is standard web infrastructure: HTTP, JSON-RPC 2.0, and streamable HTTP for live updates.\n\nThe two core primitives are AgentCards and Tasks.\n\n- An\n[AgentCard](https://agent2agent.info/docs/concepts/agentcard/)is a JSON document published at`/.well-known/agent-card.json`\n\n. It describes what the agent can do, what inputs and outputs it accepts, and how to authenticate. It's machine-readable by design, so orchestrators can discover and evaluate sub-agents programmatically. Before delegating a task, an orchestrator fetches the target agent's AgentCard at that URL to confirm it can handle the work. - A\n[Task](https://agent2agent.info/docs/concepts/task/)is the unit of work. Each task has a unique ID and moves through a defined lifecycle: submitted, working, input-required, completed, failed, canceled, or rejected. A sub-agent streams live progress back to the orchestrator as the task runs. If additional credentials are needed mid-task, the task moves to an`auth-required`\n\nstate and pauses rather than failing silently.\n\n## MCP vs. A2A? Multi-agent systems need both\n\nMCP is vertical: an agent reaches down to a tool or data source that serves it. A2A is horizontal: an agent reaches across to another agent it can delegate to.\n\nTo give you an idea of how the two fit together, here's an example workflow.\n\nAn orchestrator is building a competitor analysis. It uses A2A to find a research sub-agent and delegates the task: \"gather structured pricing and feature data for these ten companies.\" The research sub-agent accepts the task, uses MCP to call the tools it needs (web scrapers, search APIs, content extractors), assembles the results, and returns them to the orchestrator. The orchestrator never touched the tools directly; the research agent never saw the broader workflow.\n\nA fair question is whether A2A adds anything that a well-architected MCP system couldn't handle. For simple setups, not much.\n\nWhere A2A earns its place is at scale: when an organization runs many agents across different vendors and frameworks, coordinating them through individual MCP connections becomes unmanageable. A2A gives the higher-level abstraction those environments need.\n\n## What gets built on this foundation?\n\nNeither Anthropic nor Google unilaterally controls the roadmap for these protocols anymore (both were donated to the [Agentic AI Foundation](https://aaif.io) in 2025). Vendor-neutral governance is now what lets the rest of the industry confidently build on top of them.\n\nIf orchestration-based architectures become the dominant model for how agents discover tools by capability, delegate tasks without human intervention, and get results back, the services built on MCP and A2A are well-positioned to become the default infrastructure for agentic traffic.\n\nThe opening question was about how agents discover tools, call services, and coordinate across systems they don't own. These two protocols are the leading candidates for that role. What gets built on them (and whether they hold that position) is still being decided.\n\n*Read **Agentic commerce and the AI economy stack** for broader context on how MCP and A2A fit alongside x402, KYAPay, and the payment protocols shaping autonomous agent transactions.*", "url": "https://wpnews.pro/news/mcp-and-a2a-building-the-agentic-internet", "canonical_source": "https://blog.apify.com/mcp-a2a-agentic-internet/", "published_at": "2026-06-28 08:02:24+00:00", "updated_at": "2026-06-28 08:11:01.431507+00:00", "lang": "en", "topics": ["ai-agents", "ai-infrastructure", "large-language-models", "ai-tools", "ai-research"], "entities": ["Anthropic", "Google", "Model Context Protocol", "Agent-to-Agent Protocol", "AgentCard", "Task"], "alternates": {"html": "https://wpnews.pro/news/mcp-and-a2a-building-the-agentic-internet", "markdown": "https://wpnews.pro/news/mcp-and-a2a-building-the-agentic-internet.md", "text": "https://wpnews.pro/news/mcp-and-a2a-building-the-agentic-internet.txt", "jsonld": "https://wpnews.pro/news/mcp-and-a2a-building-the-agentic-internet.jsonld"}}