Large Models as Engines, Tool Ecosystems as Limbs: How AI Agents Connect Everything

1. Engine limitation: need limbs

Large models are probabilistic text‑prediction engines that can generate code, reports, and answers, but they operate only at the language layer and cannot directly access databases, call APIs, or manipulate file systems. A middle layer is required to translate intent into action.

Before MCP, developers relied on vendor‑specific Function Calling, which suffers from three systemic flaws:

Cross‑model incompatibility: OpenAI’s function format and Anthropic’s tool use are not interchangeable, forcing rewrites when switching models.

State management chaos: Function Calling is stateless, making it unsuitable for pagination, long‑running connections, or other stateful scenarios.

Redundant effort: Teams repeatedly wrap the same tools (e.g., Jira, Salesforce, GitHub), wasting resources.

Thus, a standardized “transmission system” is needed to connect the engine to its limbs.

2. MCP emerges: the USB‑C of AI

In November 2024 Anthropic open‑sourced the Model Context Protocol (MCP), an open standard that unifies communication between large language models and external data sources or tools.

Core value: write an MCP Server once and any MCP‑compatible client—Claude Code, VS Code Copilot, Cursor, ChatGPT—can invoke it without per‑platform adapters.

Three core primitives

Resources: read‑only GET‑like access exposing files, database records, or API responses to the LLM.

Tools: side‑effect operations (send email, create tickets, modify databases) driven by LLM decisions, with JSON Schema‑derived strong‑typed parameters.

Prompts: standardized interaction templates that reduce repetitive coding.

3. 2026 milestone: stateless core

On 28 July 2026 MCP released its largest revision, shifting from bidirectional stateful connections to a fully stateless core. Each request is self‑contained, carrying protocol version, client identity, and capability information; the protocol‑level handshake and Mcp-Session-Id header are removed, enabling serverless and edge deployments.

Key ecosystem data (2026): over 400 million monthly SDK downloads, 12 000+ available servers, 28 % Fortune 500 deployments, 78 % enterprise AI teams using MCP in production, TypeScript + Python SDKs exceeding 1 billion cumulative downloads, and 41 % of software organizations running MCP in production.

4. From brain to limbs: MCP in action

Resources – letting AI “see” the world

Resources let the model read external data on demand, complementing Retrieval‑Augmented Generation (RAG) and reducing reliance on training‑data memory.

Tools – letting AI “do” work

Tools enable LLM‑driven decisions to execute side‑effect actions; JSON Schema automatically derives parameter types for strong‑type validation.

Enterprise case studies

ServiceNow: Action Fabric allows external AI agents to connect via MCP to invoke workflows and playbooks.

Salesforce: Headless Data 360 supports MCP; a company launched AI assistant EVA in 12 days, handling 50 % of customer chats.

Moody’s: Integrated credit and compliance workflows into Claude via MCP for interactive reporting.

Tencent Cloud: An international courier processes over 40 complex business flows through MCP.

Major SaaS platforms (Salesforce, SAP, ServiceNow, Microsoft 365, Google Workspace, Jira, Slack) have released official MCP servers, signaling protocol maturity.

5. Beyond MCP: A2A and the “agent internet”

MCP solves “Agent‑to‑Tool” interaction; the ultimate agent form requires collective collaboration. In April 2026, Google, AWS, Microsoft, Salesforce, SAP, ServiceNow, and 150+ other organizations launched the Agent‑to‑Agent (A2A) protocol 1.0 GA.

Analogy: MCP schedules employee‑to‑tool tasks; A2A schedules employee‑to‑employee or employee‑to‑customer communication, forming the communication foundation of an “intelligent‑agent internet”.

Gartner notes that the widespread adoption of MCP and A2A marks the end of the “isolated agent” era and the beginning of a “TCP/IP moment” for AI agents.

Just as HTTP linked computers into the Internet, MCP is linking software, data, and APIs into a cyber‑network orchestrated by AI brains.

6. Conclusion

Large models provide decision intelligence; MCP‑type protocols supply execution capability. Combined, they move AI from “talking” to “doing”. By 2025 we taught AI to think; by 2026 we teach it to use tools and collaborate. The future “agent internet” depends on how we adopt these standards.