Which Nielsen heuristics work for AI agents without change?

"Augmenting Interface Usability Heuristics for Reliable Computer-Use Agents" (an arXiv preprint; the excerpt does not indicate peer review) revisits "Nielsen's ten usability heuristics (Nielsen & Molich, 1990)" to ask which hold up for "general computer-use agents" (CUAs) — "systems that operate software directly, much like humans do." The authors "identify a subset of heuristics that are directly beneficial to CUAs": consistency and standards, flexibility and efficiency of use, recognition rather than recall, and error prevention together with "mechanisms that support error diagnosis and recovery." They describe this subset as "directly transferable principles that are broadly beneficial for CUA performance," distinct from heuristics the paper says need agent-specific adaptation.

The paper's reasoning for each heuristic is interface-mechanical rather than a loose analogy to human cognition. Consistency and standards "help establish shared structures across interfaces, including common layouts, controls, and descriptions, which support knowledge reuse and improve generalization across applications" — an agent that has seen one standard dropdown can reuse that prior on an unfamiliar app. Flexibility and efficiency matter because "current CUAs still lag behind human experts in execution efficiency," so workflows with fewer required steps leave less room for compounding error. Recognition rather than recall is "highly relevant because agents rely on limited context windows to maintain interaction history, making explicit interface cues much easier to use than reconstructing prior states or procedures from memory" — an agent's working-memory limit becomes a UI design constraint. Error prevention and recovery "help guide CUAs toward correct workflows and reduce avoidable failures," paralleling their role for human users.

This reframes the Amershi-era project of distilling human-AI interaction guidelines (How should AI systems be designed across different interaction phases?) from the opposite direction: rather than writing new guidelines for how systems should treat human users, this paper asks which existing, decades-old usability heuristics for human users already serve an AI agent sitting in the user's seat and operating the interface itself. The same Nielsen whose ten heuristics anchor this analysis also supplies the vocabulary for AI-era interfaces one layer up, in How should AI interfaces handle the shift from doing to supervising? — that model addresses how a person delegates intent to an agent, while this paper addresses how the agent then executes against the resulting GUI.

The excerpt names which heuristics transfer but not which specifically require adaptation or how — that detail sits in "Table 1," which is not included here, so the non-transferable heuristics and the paper's own four new agent-specific heuristics remain unspecified in this note (see Can interface design help agents complete computer tasks better? for the augmentation results). The paper frames these four as necessary but not sufficient: heuristics that survive unmodified, not a complete account of agent-compatible design — the remaining heuristics, it says, need "safe additive augmentations" rather than replacement.