Do the classic rules of good app design even apply when the 'user' clicking around is an AI, not a person?

This explores whether Nielsen's ten classic usability rules, written for people, still work when an AI agent is the one operating the screen. The corpus gives a split answer. Some of them carry over unchanged and some need rethinking. According to Which Nielsen heuristics work for AI agents without change?, four rules transfer as-is: consistency and standards, flexibility and efficiency, recognition over recall, and error prevention. The reason is interesting. Agents have their own version of human limits. They can only hold so much context at once, and they have to work out which pixel or element on the screen corresponds to the thing they want to act on. A consistent layout helps an agent for much the same reason it helps a tired human: there's less to figure out from scratch.

The other six rules are where it gets interesting. Simply carrying over the old list isn't enough. Can interface design help agents complete computer tasks better? adds new heuristics aimed at the places agents actually struggle: perceiving what's on screen, connecting instructions to the right element, and controlling inputs reliably. Agents completed more tasks and worked more efficiently with these additions. The surprising part is that human testers saw no drop in usability. Designing for agents and designing for people don't seem to conflict, at least in this study. A screen that is clearer to a machine was not worse for a person.

A more radical view is that the best interface for an agent may not be a visual interface at all. Can API-first agents outperform UI-based agent interaction? found that letting agents call APIs directly, instead of clicking through screens step by step, cut task time by roughly two-thirds with almost no loss of accuracy. If that holds up, then usability heuristics for agents may be mostly a transitional problem. They matter while agents still have to use interfaces built for human eyes and hands. A related idea appears in Where does agent reliability actually come from?: much of what makes agents reliable comes from the scaffolding around the model (memory, reusable skills, structured protocols), not from the model itself. That is close to how Nielsen's 'recognition over recall' helps humans: you put the knowledge into the environment so nobody has to keep it in their head.

There is also a second question hiding behind this one. Heuristics can also be used to judge agents as products that people interact with. Do productivity agents actually protect users better than emotional ones? ran a heuristic-style review of nine deployed agents. General productivity tools scored well, while companion and advisory agents most often broke principles like making their intent visible and avoiding user dependency. Here the principles are about protecting the human rather than helping the agent click the right button. The corpus doesn't yet bring these two uses together, and that is a gap worth noticing.