How should AI agents balance helpfulness with staying out of the way?

Microsoft Design's cross-functional team, writing as Ruokan He, Jen Fox, and Amanda Snellinger, states that its "Agent UX Design Principles have three high-level categories with 6 sub-categories": "Connecting, not collapsing"; "Easily accessible yet occasionally invisible"; and "Embrace uncertainty but establish trust." The middle category is broken into three named sub-points — Past, Now, and Future — and the essay adds a cross-cutting statement that "transparency, control, and consistency are foundational elements of all agents," applying to the whole set rather than to one category.

Each category carries its own stated reasoning. Under "connecting, not collapsing," the authors write that agents "bring people closer together" and "are not designed to replace or belittle people," and should be "discoverable, intuitive, and accessible... on any device or platform." Under "easily accessible yet occasionally invisible," an agent draws on "memory and connected databases" to act on "richer historical data beyond singular events, user queries, or states," and rather than issuing a "static notification," it may "proactively start a chat or create artifacts" to "direct the user's attention at the right moment" — adapting across devices and modalities, for example "voice-first interaction on mobile but text-first interaction on a PC." Under "embrace uncertainty but establish trust," the authors treat uncertainty itself as "a key element of agent design," to be paired with a visible "level of certainty and reasoning," customizable settings, an explicit on/off control, and an agent status "clearly visible at all times," stated purpose being "to avoid overreliance."

The nearest comparison is What UX principles do workplace users want in AI agents?, which also names a structured set of agent UX principles meant as design guardrails. The two sets do not line up: that paper's eight principles were derived from workshops and interviews and weighted by what workplace users said they prioritize, while Microsoft's three categories are stated as the design team's own cross-functional position, not derived from a study of user priorities. Within that difference, the call to make "level of certainty and reasoning... visible and/or easily accessible to the user" extends How can users verify what an AI agent actually plans to do? from verifying an agent's plans specifically to verifying its underlying confidence more generally. The requirement that background agents carry "a user-facing mechanism to view and control actions and automations" also echoes the knowledge-and-observation half of Does human-in-the-loop design actually enable effective oversight?, though Microsoft's essay treats status visibility and control as a principle to build rather than a fallacy to diagnose.

The excerpt is a practitioner essay, not a study: it gives no methodology for how the principles were chosen beyond "we've worked cross-functionally," no evaluation of built agents against the principles, and names only three of the stated six sub-categories, leaving the sub-categories under "connecting" and "embrace uncertainty" unspecified. It does not claim the principles were tested with users or that following them changes trust or adoption. At this strength, the excerpt establishes only what one named design team at Microsoft has adopted as its own internal design principles for agents, offered as guidance rather than demonstrated result.