Agents as users
The protocol stack
The agent protocol stack is four layers with distinct jobs: Model Context Protocol connecting agents to tools and data, Agent2Agent connecting agents to each other, AG-UI streaming agent output into a first-party front end, and MCP Apps rendering interactive interfaces inside another company’s host.
Control sits with the agent
When to use it
Any integration decision, because choosing the wrong layer is expensive to undo.
What it prevents
One-off integrations that only one agent can use.
Protocol stackfour layers
AG-UIAgent output into an interface
A2AAgent to agent
MCPAgent to tools and data
IdentityUnder all three
Choosing the wrong layer is the expensive mistake.
Who ships it
- Agent2Agent and MCP Apps: A2A moved to the Linux Foundation and reached over 150 supporting organizations by April 2026, adding signed agent cards so a calling agent can verify both metadata integrity and domain ownership. MCP Apps, specification SEP-1865, became the first official MCP extension in January 2026
Sources
Related patterns: Generative interfaces · Delegated identity