Understanding the roles of agent harness, framework, and mcp in AI architecture

A recent article explores the distinctions between agent harnesses, frameworks, and MCP (Model Context Protocol) in AI systems, focusing on which layer manages execution loops, state, tools, permissions, and recovery. MCP, defined by its three server-side primitives—tools, resources, and prompts—now supports MRTR (Multi Round-Trip Requests) for handling permissions. OpenAI’s Relay and Codex exemplify how applications own product context and tools, with MCP acting as the protocol layer. The article highlights that harnesses, like Claude Code and OpenAI’s, manage permissions and recovery, while frameworks, such as LangGraph and Microsoft’s Agent Framework, provide hooks for approval logic. MCP’s role is to facilitate communication between agents and tools, with no overlap in agent-to-agent communication. The MCP specification emphasizes that security principles must be enforced by the host, not the protocol itself. Adoption figures show over 1 billion SDK downloads monthly, underscoring MCP’s significance in the agentic AI landscape. The article notes that while MCP handles transport and basic tool execution, harnesses and frameworks manage higher-level policies and recovery. OpenAI’s harness improved GPT-5.6 Sol performance, and Codex’s application-server demonstrated robust tool execution. The boundaries between these layers are blurring, but A2A (agent-to-agent) remains distinct. The piece underscores the importance of clear ownership of responsibilities to ensure secure and efficient AI agent operations.