Harnesses, frameworks and MCP divide responsibilities in agent systems
Agent harnesses, agent frameworks and the Model Context Protocol occupy different layers of an agent stack, although their functions increasingly overlap. A harness wraps a model in a working execution system: it typically owns the turn loop, sessions, context handling, permissions, sandboxing and recovery. Claude Code and Codex are cited as examples of this product-oriented approach.
A framework provides composable building blocks rather than one fixed policy. LangGraph, the OpenAI Agents SDK and Microsoft Agent Framework offer model and tool abstractions, graphs, handoffs, memory or checkpoint interfaces, guardrails and observability. Developers configure termination, persistence and approval behavior, while the framework supplies the loop structure and recovery primitives.
MCP is a protocol, not an agent runtime. It standardizes JSON-RPC 2.0 communication between hosts, clients and servers for tools, resources and prompts, using stdio or Streamable HTTP. It owns tool transport but does not own the agent loop, core state or permissions; the host must obtain user consent and enforce security. An MCP tasks extension can support long-running tool calls, but agent-level recovery remains the responsibility of the runtime. The article presents the layers as complementary: the application sets business rules, the runtime executes work, and MCP connects capabilities.
Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.