What's the difference between an agent harness and an agent framework?

An agent framework is a library you build an agent with, such as CrewAI or the OpenAI Agents SDK; an agent harness is the running system around the model in one agent: its prompts, tools, context handling, hooks, sandbox, and the loop that runs them. Every agent has a harness, and a framework is one optional way to build it. Addy Osmani’s write-up defines a harness as “every piece of code, configuration, and execution logic that isn’t the model itself,” so a framework running your agent is part of its harness too.

The best-known coding harnesses use no framework. A source-code study of eleven coding harnesses, including Claude Code, Codex CLI, and Aider, found that “no agent runtime imports a general-purpose agentic framework.” The Claude Agent SDK blurs the line: it’s a library that ships Claude Code’s finished loop rather than building blocks. LangChain uses harness for that kind of library, calling LangChain a framework, LangGraph a runtime, and Deep Agents, which adds default prompts, planning tools, and a filesystem, a harness.

So a framework upgrade is a harness change even when none of your own code moved, and it can shift results just as a prompt edit can.