What new design rules are people proposing just for AI agents that click and type their way through software?

This explores the design rules being written for AI agents that operate computers through the same screens, buttons and forms people use. The interesting part is that this work runs backwards from the usual direction. It doesn't train the agent to cope better with messy interfaces. It asks whether the interface itself can be built to help the agent.

The starting point is that a good part of the old rulebook already works. Of Jakob Nielsen's ten classic usability heuristics, four carry over to agents unchanged: consistency and standards, flexibility and efficiency, recognition rather than recall, and error prevention Which Nielsen heuristics work for AI agents without change?. They transfer because agents struggle in ways that resemble human struggles. They can only hold so much in view at once, and they have to work out which element on the screen matches what they're trying to do. An interface that doesn't make a person remember hidden state also doesn't make an agent reconstruct it.

The new heuristics cover where agents differ from people. The UI-Verse work groups its additions under three agent problems: perception (can the agent tell what's on the screen?), grounding (can it tie an instruction to the right element?) and control (can it reliably carry out the click, scroll or entry?) Can interface design help agents complete computer tasks better?. Interfaces redesigned along these lines raised agents' task completion and efficiency. The surprising result is that human users showed no observable drop in usability. That suggests designing for agents doesn't have to mean building a second, machine-only web. The available summary names the three categories but doesn't list each rule, so for the specific heuristics, open that paper directly.

Two neighboring ideas sit near this. One moves the effort from the interface to the agent's surrounding software. Automated searches across many environments found four mechanisms in the agent's wrapper code that cut token use by nearly half: how actions are executed, how context is compressed, how observations are handled, and how reading is delegated Can agent harnesses be automatically optimized across many environments?. That is the same perception-and-control problem, solved from the agent's side. The other turns heuristics into a way of judging agents rather than designing for them. An evaluation of nine deployed agents against 14 HCI principles found that companion and advisory agents most often violated the high-risk principles, such as preventing user dependency and making the agent's intent visible Do productivity agents actually protect users better than emotional ones?.

The takeaway is that the screen is becoming a shared workspace for two kinds of user, and the evidence so far says you can serve both at once. The corpus is still thin here, though. The only work that proposes new agent-specific heuristics is UI-Verse, so treat this as an early design agenda, not a settled checklist.