AI agents change the main question in software design

For thirty years, the person at the keyboard was the one who acted. Software waited for a click or a keystroke, then did exactly what the person asked. Interface design grew up around that fact. Its job was to make the right action clear and easy, because a person was going to press the button and live with the result.

That is changing: people now hand work to AI agents. A person describes the result they want, and the agent decides which tools to use, which data to open and in what order. Often no person reviews any single step. The click is still there, but it now approves a whole series of decisions instead of one.

This changes the main question designers must answer. The old question was “can a person use it correctly?” That question assumes the person does the acting, and it is about the screen and its buttons. The new question is “can we stop the agent doing something the person never intended, with the tools and data it has been given?” That question assumes the agent does the acting, and it is about the limits of the whole system.

The two kinds of failure are different. When a screen is confusing, a person wastes a few minutes and the problem is fixed in the next release. When an agent is not properly limited, it may send protected data somewhere it was never meant to go, or act beyond its authority when it receives an input nobody tested. That failure happens once, with no warning, and the damage lands wherever the data or the authority ended up.

Kern, Syntaxe’s security product for AI agents, is built to solve this problem. It does not watch an agent’s behaviour and raise an alarm when something looks wrong. Instead, it takes the tools each agent is allowed to use, the labels on the data and the limits on each agent’s authority, and turns them into a precise model of the system. It then proves, for every possible input, what the agents can and cannot reach.

Software that is easy to use is still worth having. But any software that acts on your behalf now needs clear, proven limits before it is released.