Agent Harness, Framework, MCP Differentiated in AI Architecture

Discussions surrounding AI agent architecture often use the terms "agent harness," "agent framework," and "MCP" interchangeably, yet they represent distinct layers with specific responsibilities. Understanding these differences is crucial for building robust and predictable AI systems. The "agent harness" functions as the core execution system, wrapping a large language model (LLM) to create a functional agent. OpenAI's Codex platform, in an August 19, 2026, post, defined the harness as managing conversation state, streaming execution, and facilitating tool usage. It also enforces sandbox environments and approval policies, carrying work across multiple conversational turns. Anthropic's Claude Code documentation refers to this same concept as an "agentic harness," with the Claude Agent SDK providing access to the tools, agent loop, and context management that power Claude Code. An agent harness is inherently opinionated, delivering a complete package that includes the execution loop, a permission model, a sandbox, and a context strategy as a single unit.

In contrast, an "agent framework" serves as a library of fundamental components for composing agents. It provides abstractions for model clients, tools, graph orchestration, memory interfaces, and observability hooks. Prominent examples of agent frameworks include LangGraph, the OpenAI Agents SDK, and Microsoft Agent Framework, which achieved General Availability (GA) 1.0 in April 2026. An agent framework offers the building blocks and a skeletal structure for the execution loop, allowing developers to define their own policies and specific implementations. This approach provides flexibility, enabling users to customize agent behavior beyond the defaults offered by a harness.

The "Model Context Protocol" (MCP), however, operates at a different level entirely. MCP is not a runtime system but rather a wire protocol designed to standardize how an LLM application, referred to as the host, discovers and interacts with capabilities exposed by servers. These capabilities can include tools, resources, and prompts. MCP utilizes JSON-RPC 2.0 messages for communication between hosts, clients, and servers. Since December 2025, the Linux Foundation's Agentic AI Foundation has been the governing body for MCP, overseeing it alongside other initiatives like goose, AGENTS.md, and A2A. Crucially, MCP does not own the execution loop or any agent state; its primary function is to define and manage the contract between an agent and its tools, ensuring interoperability and clear communication channels.

An ownership matrix clarifies the default responsibilities of each layer. The agent harness owns the execution loop, conversation state, tool transport, permission enforcement, and recovery mechanisms. The agent framework owns the composition of tools, the definition of the loop skeleton, and the integration of memory and observability. MCP, as a protocol, owns the standardized communication contract for tool discovery and invocation. While these layers have distinct primary responsibilities, their functionalities increasingly overlap at the edges, requiring careful consideration during the design and implementation of complex AI agent systems. The distinction is vital for developers aiming to leverage the strengths of each component for specific AI applications.