How can people delegate work while retaining meaningful control?
Here, agentic AI means systems that use tools to take steps towards a goal, potentially changing files, sending messages or acting in other services. The UX includes setting boundaries, understanding progress, intervening and recovering—not only writing a prompt.
5 supporting readings · Sources checked 18 October 2026
Design for delegation, not blind reliance
Boundaries: distinguish proposing, drafting and executing. Make scope, permissions and consequential actions understandable.
Visibility: show completed actions, pending decisions, uncertainty and failures. A plausible explanation is not proof that an action succeeded.
Intervention: provide usable pause, stop, correction and handover controls. Avoid asking for approvals so routinely that they become meaningless.
Recovery: show what changed and what can be undone. Do not imply that an irreversible action has an undo option.
Evaluation: examine task success, unauthorised actions, error detection, recovery and oversight effort. Treat appropriate reliance as the goal, rather than maximum trust.
Editorial questions informed by the readings below; they are not a validated checklist or a compliance certification.
Papers and supporting evidence 5 items
Three peer-reviewed background papers, one industry empirical report and one practitioner guide. Older publications retain their dates. Peer review, conceptual argument and observational evidence answer different questions.
Saleema Amershi and colleagues · CHI 2019 · 2019 Open author manuscript · full text reviewed
A set of 18 guidelines, refined through several evaluation stages, including 49 design practitioners assessing 20 AI-infused products. Useful for examining expectations, correction and user control across an interaction.
Limitations: Guideline evaluation is not evidence that a particular interface improves user outcomes. The products studied predate current LLM agents; applicability needs testing in the intended context.
Consider: Where can someone discover the system’s limits, correct a mistake and change its behaviour?
Eric Horvitz · CHI 1999 · 1999 Open author manuscript · full text reviewed
Examines how automated assistance and direct user control can work together, illustrated by the LookOut scheduling prototype. It connects intervention timing and uncertainty about a user’s intentions to the costs of acting.
Limitations: A principles and prototype paper, not a modern LLM-agent trial or a general estimate of effectiveness. Its design arguments require contextual evaluation.
Consider: When should the system act, ask a question or leave the task to the person?
Lisanne Bainbridge · Automatica, 19(6), 775–779 · November 1983 Abstract-only review · publisher full text may require access
The abstract argues that automating industrial processes can expand, rather than remove, problems involving the human operator. A historical starting point for questioning what remains for people to do.
Limitations: Only the abstract and bibliographic information were available for this review. No detailed findings or effect sizes are claimed. Industrial process control is not equivalent to contemporary agentic AI.
Consider: What difficult work remains for the person after the routine work is automated?
Anthropic · Anthropic research report · 18 February 2026 Open report · methods and limitations reviewed
An observational analysis of millions of interactions across Claude Code and API usage. More experienced Claude Code users both enabled more automatic approvals and interrupted more often, suggesting oversight is more complicated than approval frequency alone.
Limitations: Vendor-authored evidence from its own ecosystem. Changing users, tasks and product features complicate causal interpretation; turn duration is an imperfect proxy for autonomy. This does not establish that more autonomy is safer.
Consider: How would you distinguish useful intervention from unnecessary approval friction?
Microsoft · Microsoft Learn · Living guidance · checked 18 October 2026 Public page text reviewed · some linked material may require sign-in
A starting point for agent design covering responsible AI, accessibility and interaction foundations. Use it to locate implementation guidance and identify questions for your own evaluation.
Limitations: Platform-provider guidance is not independent validation or evidence of user outcomes. Publication date was not established; recommendations may change.
Consider: Which recommendation can you turn into an observable test of user understanding?
Editorial learning exercise · allow 60–90 minutes for the first prototype; evaluation takes additional time.
Design a fictional conference-planning assistant that can compare events and draft a plan. It must request approval before booking, paying or sending a message. Use synthetic information; no live accounts or transactions are needed.
Define the agreement. Map allowed actions, approval points and forbidden actions. Explain who is affected and what happens when the brief is ambiguous.
Prototype the difficult moments. Show a proposed action, progress, an unexpected change, a pause and a failed action. Include an action that cannot be undone. Make status messages and controls usable with a keyboard and understandable without colour.
Evaluate and revise. After any required ethics approval, ask participants to explain what they authorised, identify an error and stop or recover the task. Record misunderstandings, recovery success and oversight effort. A small formative study identifies issues; it does not establish general safety.
Portfolio evidence: a permission map, an annotated flow, a short evaluation plan, findings and one documented revision. Link each design decision to a source or observation, and state what remains untested.
For more advanced work, compare approval before every action with approvals at consequential decision points. Define comparable tasks and record both missed errors and unnecessary interruptions; do not assume either pattern is superior.