AI Harnesses: The Key to Useful AI

For the past couple of years, the news about AI has focused on this or that new AI model. But AI models aren’t the only part of an AI system that matters, just like a car engine isn’t the only part of a car that makes it useful. All AI models that everyday users work with are part of some type of system. At first, that system was just a chatbot, and it didn’t really add much that users could see. This year, Claude Cowork changed that by offering a suite of tools that manage connections, spin up behind-the-scenes agents, and work directly with folders and files on the user’s desktop.

Claude Cowork, alongside Claude Code, belongs to a class of software known as a “harness,” and we’re seeing more of these as the limits of chatbots become clearer. Harnesses provide much more control over processes and over the content that an AI model is given access to. They also make it much easier to create and use a new type of file, called a “Skill.” These are basically instructions you can use over and over again, as if your favorite prompt were saved so you could access it again and improve it over time.

For those of you who have heard of OpenClaw, it is also a harness. Recently, Microsoft licensed Claude’s Cowork product, which is currently in limited availability but will eventually be available to all Copilot users.

The takeaway: useful AI systems need tools and means of control, and harnesses are the best system type out there for this. If you are a Claude user, download the desktop app and try it out. If you’re a ChatGPT user, their Codex desktop is becoming a harness for non-developers; it’s just not quite there. For Copilot, as always, give Microsoft some time before you take it for a spin.