Agent Frameworks and Harnesses: Stop Comparing a Toolbox to a Whole Car

Shubham Saboo posted on X, and SuperAGI reposted it. The content is not complicated, but it makes clear a distinction that is often conflated: AI Agent frameworks and harnesses are not the same kind of thing.

A framework is a library that developers import into their code. ADK, LangGraph, and CrewAI provide parts such as agent classes, tool interfaces, and model adapters. How the loop runs, what goes into the context, and how the agent checks its own work are still up to the developer.

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

The author's choices are direct:

• If the value lies in the domain, data, or workflow, start with a harness.
• If the loop itself is the product, or you need to control and audit every step, use a framework.
• If you need both, mix them. Give open-ended tasks to the harness and connect a framework where determinism is needed.

The boundary is disappearing. LangChain used to be a framework company, and now it has Deep Agents, a harness; Claude Code, a harness, can also be imported as a library through the Claude Agent SDK. If teams still frame the choice as "framework or harness," they are likely to argue about the wrong thing.

One judgment: when choosing, first look at where your value lies. If data, processes, and domain knowledge are the moat, a harness saves the cost of rebuilding the loop. If every step must be controllable and auditable, and the loop itself determines the product experience, a framework is more suitable. The key to mixing them is division of labor: give open-ended work to the harness and connect deterministic work to the framework.

Comparing the performance of a framework and a harness is like comparing a toolbox and a whole car for acceleration. First confirm whether you are buying parts or a car that can get on the road.