How should agents decide when to act versus ask users?

Horvitz's 1999 CHI paper "Principles of Mixed-Initiative User Interfaces" sets out twelve principles for interfaces where "intelligent services and users... collaborate efficiently to achieve the user's goals," rather than choosing between direct manipulation and full automation. The paper frames the debate it answers: one group of researchers pursued "interface agents" that sense activity and take automated action, another pursued better direct-manipulation metaphors, and Horvitz argues the two should be coupled rather than treated as competing tacks. The key problems he names with agent-based interfaces are "poor guessing about the goals and needs of users, inadequate consideration of the costs and benefits of automated action, poor timing of action," and inadequate means for users to guide or correct automated results.

The twelve principles build on a few recurring mechanisms rather than standing as a flat checklist. Systems should infer a user's goals probabilistically under uncertainty, "employing machinery for inferring and exploiting the uncertainty about a user's intentions"; time action against a model of the user's attention, deferring when "action will be less distracting"; weigh the expected costs and benefits of a service before invoking it; use dialog to resolve uncertainty "considering the costs of potentially bothering a user needlessly"; give users efficient means to directly invoke or terminate services; minimize the cost of wrong guesses through graceful timeouts and easy rejection; scope service precision down under uncertainty rather than guess boldly, preferring "doing less" but correctly; support collaborative refinement of agent output; use socially appropriate default behaviors; maintain working memory of recent shared interactions; and keep learning from observation. Horvitz illustrates the principles with LookOut, a scheduling and meeting-management agent built into email.

Since When should human-agent systems ask for human help? names interrupt-timing as its "fundamental challenge" for a modern multi-agent orchestrator, Horvitz's framework supplies the earlier articulation of the same problem: expected cost-benefit reasoning under goal uncertainty, in place of a ground-truth label, as the basis for deciding when to act versus defer. What UX principles do workplace users want in AI agents? reports which of these values workplace users say they want today; Horvitz's principles are the designer-facing mechanisms — uncertainty modeling, dialog, direct override, graceful degradation — that aim at that same control and reliability two decades before the survey measured demand for it.

The excerpt is drawn from the abstract, introduction, principles list, and conclusion, and does not include LookOut's evaluation data, so it supports the principles as design guidance Horvitz offers, not as a validated outcome — no measurement of reduced interruption cost or improved task completion appears here. It also predates today's LLM-based agents by two decades: the inference machinery it describes is Bayesian-network goal inference, not language-model reasoning. The principles should be read as a framework later agentic-interface work builds on and departs from, not as a description of how current agents are built.