Agentic Protocols

This note tracks two important interoperability protocols in the current agent ecosystem: MCP for agent-to-tool integration and A2A for agent-to-agent delegation. Both are active reference points, but they are not equally mature in adoption or tooling support.

Maturity and Scope

MCP has broad mindshare and a growing implementation ecosystem across model hosts, IDEs, and tool adapters.
A2A is a newer protocol proposal focused on inter-agent collaboration; its conventions are useful to track, but deployment patterns and governance are still settling.
This note focuses on MCP and A2A because they cover two distinct boundaries in multi-agent systems. It does not claim they are the only interoperability patterns in use.

Model Context Protocol (MCP)

Developed by Anthropic and adopted across a growing set of tools, MCP is an open protocol for connecting model hosts to external data sources and executable capabilities.

Host/Client/Server Model: Decouples the AI application (Host) from the capability provider (Server).
Layered Design: Separates the message semantics (Data Layer) from the communication channel (Transport Layer).
Primitives: Standardizes how Tools, Resources, and Prompts are discovered and used.
Client Capabilities: Enables servers to leverage the host's LLM (Sampling) or interact with the user (Elicitation).

For a deep dive, see mcp-architecture.

Agent-to-Agent (A2A)

A2A is an agent-to-agent protocol effort associated with Google and broader community work on inter-agent communication. In this vault, it is treated as a peer-to-peer complement to MCP: where MCP handles agent-to-tool communication, A2A models delegation and multi-turn collaboration between agents with their own internal state and control logic.

Agent Card / Skill Model: An agent may publish an Agent Card describing identity, endpoints, authentication requirements, and advertised Skills. In current materials, the common discovery path is /.well-known/agent-card.json.
Stateful Task Model: Interaction is framed around a Task object with a lifecycle such as SUBMITTED → WORKING → COMPLETED, plus states for failure or additional input.
Transport Flexibility: Current A2A materials describe request/response, streaming, and async notification patterns for longer-running work.

For the full technical model, see a2a-protocol. For the complementarity relationship with MCP, see a2a-mcp-contrast.

Limitations and Context

These protocols address different boundaries in a system, but real deployments may still rely on vendor APIs, framework-specific adapters, or custom orchestration layers.
MCP is farther along in practical tooling today. A2A is more useful as a design and interoperability reference than as a universally deployed default.
If a workflow only needs tool use inside a single host, MCP may be sufficient; A2A matters when distinct agents must advertise capabilities, exchange task state, or hand work across trust boundaries.

References