AIAcademy · AIAcademy · 2026-05-16
A year ago, "how does an agent talk to a tool" was an engineering question with twelve vendor-specific answers. It's now a three-layer stack, and the layers don't compete — they compose.
MCP is the tool layer. Model Context Protocol, open-sourced by Anthropic in November 2024, is the boring part that won: a shared client/server protocol, an official registry, and native support across major agent clients. The 2026 MCP roadmap is now about audit trails, SSO, transport scalability, and gateways — the unglamorous enterprise-readiness work that signals a protocol has arrived. MCP is to agents what HTTP was to the web.
A2A is the coordination layer. Google's Agent-to-Agent protocol grew from 50 launch partners in April 2025 to 150+ a year later, with a public-RFC governance model and v1.2 shipped. Where MCP lets one agent call many tools, A2A lets many agents call each other across vendor boundaries. Microsoft Copilot Studio, Salesforce Agentforce, and most multi-agent orchestrators now speak it natively.
Skills is the packaging layer. Agent Skills, introduced by Anthropic in October 2025 and opened as a standard in December, defines a SKILL.md folder format for procedural knowledge an agent can load on demand. Microsoft VS Code, GitHub, Cursor, Goose, Amp, and OpenCode adopted it directly; OpenAI's Codex CLI implements it structurally. Skills is to agents what npm was to JavaScript: a unit of reusable behaviour.
The mental model: MCP exposes capabilities, A2A coordinates agents, Skills packages knowledge. Anyone teaching agent-building in 2026 needs all three on the board. Anyone still pitching their proprietary tool-calling spec is selling lock-in.