If an AI is the one clicking buttons on your app, do your usual design rules for humans still apply?

This explores which classic usability rules, such as showing what the system is doing, keeping the user in control, and preventing errors, have to be rethought once an AI agent is either using the interface or acting as the interface. The corpus treats this as two separate problems. One is designing screens that agents can operate. The other is designing agents that people can operate.

The most direct evidence is about the first problem. In the UI-Verse benchmark, researchers added interface heuristics aimed at three things agents struggle with: perception (can it see what's on screen), grounding (can it match a label to the right clickable element), and control (can it reliably act on that element). The adapted designs consistently raised agent task completion and efficiency. Human studies also found no drop in usability for people Can interface design help agents complete computer tasks better?. So designing for agents doesn't have to cost anything for humans. Interfaces can serve both audiences at once, which weakens the idea that we'll need separate 'agent web' and 'human web' versions.

The second problem breaks some older assumptions. Classic heuristics like 'consistency' and 'recognition over recall' assume a stable interface that users can learn. AI context isn't stable. It is made of prompts, conversation history, retrieved data, and hidden state that keep shifting, so users can't form a reliable mental model of it How does AI context differ from conventional software context?. Consistency then has to come from somewhere else: you engineer what the agent knows rather than polishing what the screen shows. Fin's team puts this into practice by keeping factual content in retrievable knowledge bases and reserving agent instructions for tone and behavior Should factual content and behavior rules live separately?. They also show that conversation design decisions, such as tone, level of detail, and when to hand off to a human, have measurable effects. A single warm opening message raised customer satisfaction by more than five points Does conversation design ownership measurably improve AI agent performance?.

'User control and freedom' also needs a new form. In chat, control usually means rewriting your prompt after something has already gone wrong. Cocoa shows a stronger version: users split a plan into steps, assign each step to themselves or the agent, and revise the plan while it runs. This gave users more steering power than chat without making the system harder to use Can users steer AI agents better by revising plans mid-execution?. Error prevention changes too. Agents that use tools can drift away from what the user meant by quietly chaining actions. Conversation analysis gives a formal account of when an agent should pause and ask a clarifying question instead of searching on, which works as an error-prevention rule for agents When should AI agents ask users instead of just searching?. The difficulty is that LLM agents are passive by design and seldom take the initiative to ask Why can't conversational AI agents take the initiative?. Because they don't ask on their own, this behavior has to be designed in deliberately.

One result you might not expect: the most basic heuristic, matching the interaction style to the task, may be the one AI products break most often. Medeiros argues that defaulting to chat is a step backward for structured work. Typing commands brings back the extra layer that direct manipulation (clicking and dragging) was invented to remove. Medeiros proposes two gate questions, how fast the task needs to be and where its value comes from, to decide whether a task should be conversational or clickable Is chat the wrong default interface for AI products?. The corpus has no unified, heuristic-by-heuristic revision of a framework like Nielsen's ten. What it offers is evidence for specific adaptations: visibility, consistency, control, and error prevention each change shape once an agent is involved.