Agent Harness vs Agent Framework vs MCP: Which Layer Owns the Loop, State, Tools, Permissions, and Recovery

Agent harnesses, frameworks, and the Model Context Protocol (MCP) serve distinct roles in AI agent architecture, differing in ownership of execution loops, state management, tool transport, permissions, and recovery. Harnesses provide fixed loops and state control, frameworks offer modular primitives and loop skeletons, while MCP standardizes communication protocols without managing loops or state.