Agent Protocol Stack
The set of open standards that let AI agents talk to tools, to data, and to each other — MCP for tools and resources, A2A for agent-to-agent collaboration, plus the supporting protocols around them.
Production AI is not a prompt. It is a system of context, tools, permissions, traces, evals, and feedback loops.
What it is
The agent protocol stack is the analogue of HTTP/TCP/IP for the AI agent era: a set of layered open standards that let agents discover each other, expose tools, exchange messages, and share context without each integration being bespoke. The two most-developed pieces today are MCP (Model Context Protocol, Anthropic, 2024) for agent-to-tool, and A2A (Google, 2025) for agent-to-agent.
Why it matters
Without shared protocols, every agent integration is custom and every vendor lock-in is permanent. With shared protocols, the same MCP tool server can be used by Claude Desktop, Cursor, ChatGPT, and an internal agent — and replacing one of those clients with another doesn't require rewriting the tools. Standards are how an ecosystem becomes durable.
What's in it today
MCP (tools, resources, prompts), A2A (agent-to-agent), supporting work on authentication and trust (OAuth flavors tuned for agent identity), and emerging conventions around long-running task lifecycles. The stack is early — expect material change over the next 12-18 months.
Related resources
Anthropic's open standard for exposing tools, resources, and prompts to AI models — released in late 2024, broadly adopted across the agent ecosystem, the connective tissue of modern AI tool integration.
A protocol for agents to discover, describe, and call each other directly — Google's open spec for letting independent agent systems collaborate without each one becoming a tool for the other.
A structured description of what an AI agent can do, how to talk to it, and what trust it requires — the public face of an A2A-compatible agent.
A discoverable catalog of agents — internal and external — with their Agent Cards, capabilities, trust scope, and operational health, so other agents and humans can find and call them.
The mechanism by which a language model invokes external functions — APIs, databases, code execution, retrieval — and reads the results back to continue its work.