Background AI Agent UX: Designing for Tasks That Run While Users Are Away
How to design long-running AI agent tasks: starting them, showing status, notifying at the right moments, pausing for input and presenting results.
Quick answer
Background agents work on tasks while users do something else: researching, processing batches, monitoring, preparing drafts. Their UX has five moments: start (scope and limits), status (what is happening, without demanding attention), interrupt (asking for input at the right time), result (what was done, with evidence and partial outcomes) and steer (pause, change, cancel).
The core rule: make progress visible but quiet, and make interruptions rare, clear and actionable.
Why long-running tasks need different UX
Chat assumes the user is present and waiting. Background agents break that assumption. They may run for hours, wait for approvals, retry failed steps, change plans and finish with partial results. Engineering handles survival across crashes and waits (see durable execution for AI agents); UX handles how people delegate, follow and receive that work. Short waits are covered by the progress patterns in AI UX design; this article is about work measured in minutes to days, typical of agentic workflow automation.
Starting a task
Capture enough to avoid interruptions later: the goal, scope, limits (budget, recipients, systems), what counts as done, and decisions the agent may take alone versus those it must ask about. Show a short plan and an honest idea of duration if it is predictable. Offer to notify on completion and let the user choose channels.
Showing status without demanding attention
Put background tasks in a task list or inbox rather than in the chat scroll. Each task card shows the goal, status from a small set (queued, running, waiting for you, blocked, done, failed, cancelled), the current step in plain language, progress if meaningful, and the time of the last update. Avoid fake precision: an estimate that keeps slipping is worse than none.
start ─▶ queued ─▶ running ──────────────▶ done
│ ▲ (results +
│ │ answer receipt)
▼ │
waiting_for_you ── no answer by deadline
│ └─▶ safe default
▼ (skip step / stop)
blocked ─▶ notify + suggested fix
│
failed ──▶ partial results + retry option
user can: pause · edit scope · cancel at any pointNotifications and interruptions
| Event | Notify? | How |
|---|---|---|
| Needs a decision | Yes | Specific question, options, deadline, safe default |
| Blocked or failed | Yes | What failed, impact, suggested next step |
| Finished | Yes (if the user opted in) | Summary, results link, anything needing review |
| Important time-sensitive finding | Yes | What was found and why it matters now |
| Routine progress | No | Visible in the task card only |
Pro tip
Batch questions. If the agent will need three decisions, ask them together at the first point it must stop, rather than interrupting three times.
Pausing for input safely
When the agent needs a person, it should pause only the dependent step and continue independent work. The request should stand alone: what is needed, why, the options with consequences and when it will apply the safe default. The safe default for consequential actions is to not act. See AI action confirmation UX for which actions need approval.
Presenting results
A finished task should open on a summary: what was achieved, what was not and why, items needing review, and links to evidence and the full activity log. Partial success is normal for agent work; present it honestly ('Processed 182 of 200 invoices; 18 need your review') rather than as a failure or a success. Every action taken should have a receipt and, where possible, an undo; see AI agent trust UX.
Steering and cancelling
Users should be able to pause, change the scope or limits, and cancel at any time. Cancelling should say what has already happened and what will be reversed, since some completed steps (emails sent, orders placed) cannot be undone automatically. Handing the task to a colleague should carry its full context; see AI agent handoffs.
Design checklist
- Capture goal, scope, limits and done-criteria at the start
- Show tasks in a list or inbox, not only in chat
- Use a small set of clear statuses with plain-language current step
- Notify only for decisions, blocks, completion and urgent findings
- Batch questions; give deadlines and safe defaults
- Present partial results honestly with items to review
- Offer pause, edit and cancel; explain what cancelling cannot undo
- Keep an activity log and receipts for everything the agent did
Building agents that run long tasks?
ZSpace Labs designs task inboxes, notifications and review flows for background agents, backed by durable workflows. See AI automation and UI/UX design.
Conclusion
Background agents change the interaction from a conversation to delegation. Design how work is handed over, how status stays visible without noise, when and how the agent interrupts, how partial results are presented and how users steer. Done well, people can delegate real work and come back to clear outcomes. For the full pattern set, see AI interface patterns.
Common questions.
An agent that works on a task for minutes, hours or days without the user watching, such as researching suppliers, processing a batch of documents or monitoring orders, and reports back when it needs input or has finished.