A2A (Agent-to-Agent) and MCP (Model Context Protocol) can be layered to build extensible multi-agent MLOps systems. A2A provides the communication bus for agent discovery and coordination, while MCP acts as a universal interface for accessing tools and resources. The architecture decouples orchestration logic from execution logic: an Orchestrator agent breaks high-level goals into tasks and delegates to specialist agents (Validation, Deployment), which discover and invoke MCP tools dynamically. This pattern enables adding new capabilities without rewriting core logic, moving from rigid pipelines to adaptive, goal-oriented automation. The approach applies beyond MLOps to any domain requiring dynamic agent collaboration.
Table of contents
IntroductionLayered Protocols in Our WorkflowMCP: The Domain-Specific LanguageMLOps WorkflowSequence Diagram for the workflow:Execution and flowCode Pattern WalkthroughMCP ServerMCP ClientExecution Helpers for AgentsOrchestrator AgentValidation AgentDeployment AgentScript to start all the AgentsArchitectural Benefits of Layering those two protocolsConclusionAbout the Authors201 Impressions