Harness.
Also known as: Agent harness, AI harness
The software that wraps around an AI model and helps turn it into something you can use. Think of the model as the engine and the harness as the rest of the car. The harness can manage the instructions and context the model receives, the tools it can use, and the permissions around what it can do. An agent will typically have a model and a harness around it, and some systems use more than one model. The same model in a different harness can behave like a different product.
In the wild
“the integration between model and harness is where true agent differentiation is found”
Seen here: Ben Thompson, Stratechery, 2026
For AI users
When you compare AI tools, don’t judge them only by which model they use. Ask what each tool can connect to, what information it can reach, what it remembers, and what controls sit around it. Two products using the same model can give you a different experience.
For business leaders
The harness should be designed around the work you’re trying to do. A better model won’t fix a badly designed workflow. Some harnesses can route different tasks to different models, or be configured to switch models if one is unavailable or changes. That can help with cost and continuity, but it is not automatic in every system. A replacement model needs testing: it may behave differently, need different instructions, or fall under different data and contract terms. Run evals on the whole workflow, not just the model. Test whether it completes the job, respects its permissions, handles failures, and produces work your team can use.