NN/g: what is an agent harness and why designers need to know
As AI agents move from demos into shipped products, the vocabulary around them has started to settle. One concept that product and experience designers increasingly encounter is the agent harness — the software layer that sits between an AI model and the tools it uses.
What the article covers
Nielsen Norman Group’s Caleb Sponheim published this November 2026 article to give designers a working definition they can use in product conversations. The article is aimed at anyone who designs products where an AI agent takes multi-step actions on a user’s behalf.
What an agent harness does
The harness is the code that runs the execution loop of an AI agent. When a model decides to call a tool — checking a calendar, reading a file, sending a message — it is the harness that actually executes that call, checks whether the action is permitted, returns the result to the model, and moves the loop forward. Without a harness, a model can only reply to prompts; with one, it can work through a sequence of tasks.
The harness does three things designers need to understand. First, it assembles context: before the model sees a user’s request, the harness gathers prior messages, saved preferences, connected data sources, and system instructions. What the model can do depends partly on what the harness passes to it. Second, it connects tools and enforces permissions: the harness controls which tools the agent may call and when it must pause for the user’s approval. Third, it sets limits: the harness can cap the number of steps, the cost of a run, or the time it takes, and can halt the loop if something goes wrong.
Why it matters for designers
The practical implication is that the same underlying AI model can behave very differently depending on the harness around it. A product that uses Claude or GPT-4o as the model is not defined solely by those models’ capabilities — it is defined by what the harness exposes.
Sponheim draws out three UX consequences. First, when evaluating what a product can do, check the harness, not just the model: a task a model can theoretically handle may be inaccessible if the harness does not connect the right tools. Second, the design of an agent-powered product includes the harness behavior: permission dialogs, confirmations, preference storage, and cancellation flows are all part of the experience even if they are written by engineers. Third, in AI evaluations, test full tasks including failure cases, and work with engineering to identify whether a failure originates in the interface, the harness, or the model — the fix depends on which layer failed.
Who it is useful for
The article is useful for product designers working on AI-assisted features, design engineers, and product managers who need a shared language with engineering teams. It assumes some familiarity with AI products but requires no technical background.