Agent framework and harness: the difference isn't in the name

Agent teams often argue about one question: should we use a framework or a harness?

Shubham Saboo broke down the two terms on X. A framework is a library you import to build an agent. ADK, LangGraph, and CrewAI provide the parts: agent classes, tool interfaces, and model adapters. How the loop runs, what goes into context, and how the agent checks its own work are still up to you.

A harness is an agent that can already run. Claude Code, Codex, Deep Agents, and Hermes come with a loop, tools, context, memory, and subagents. What you bring in are instructions, skills, tools, and data.

The trouble is that the boundary is getting blurrier. LangChain used to be a framework company; now it has also launched its own harness, called Deep Agents. And Claude Code, a harness, can be imported like a library through the Claude Agent SDK. In the replies, Blum put it directly: the real confusion is that frameworks are starting to ship harnesses, and harnesses are turning into libraries.

On how to choose, Shubham gave three points: when the value is in the domain, data, or workflow, start with a harness; when the loop itself is the product, or you need to control and audit every step, use a framework; if you need both, mix them. Let the harness handle open-ended work, and bring in the framework when you need determinism.

The replies added quite a few details. Vignesh Palaniappan mentioned that an out-of-the-box harness, with sandbox permissions and tools configured, can also produce deterministic behavior. Shubham replied that in most real agent work, the value is in the tools and data, not in the loop itself, and a well-configured out-of-the-box harness is simpler and more reliable; when you need ironclad determinism or to audit every decision node, a framework or a hybrid approach is more solid.

Varun Chopra offered a test: if you keep fighting the harness's loop and want it to run your way, you probably need a framework; if you're rebuilding memory and subagents from scratch, you probably need a harness. Mladen Lotar thinks most teams will end up mixing the two: start with a harness, and once agents do strange things in production, they'll want their own loop again. Aman put it more briefly: a framework gives parts, a harness gives an opinion, and most teams need an opinion. Fahadh Kasala added one point: harnesses contain the most real bugs.

Ordinary users don't need to memorize product names. Remember one thing: who decides the loop, who manages context, and who handles memory and subagents. A framework leaves the decisions to you; a harness makes a set of decisions for you first. For people with a development background, the question to judge is whether your product value lies in the loop, or in domain data and the toolchain.

Video accompanying the original tweet: