Key takeaways
- Everyone is shipping AI agents; almost no one is designing the stop button — that's the biggest UX gap right now.
- Make the stop control loud, persistent and reachable by keyboard and screen reader at every step.
- Preview consequential actions and pause before spending money, contacting people or deleting data.
- Flag irreversible actions clearly before they run, and offer undo wherever possible.
- Control, not autonomy, drives adoption — users let agents do more when they trust they can intervene instantly.
Everyone is racing to ship AI agents. Few teams are designing the stop button. That gap is the single biggest UX risk in the current wave of autonomous features — and it echoes a lesson we should have learned decades ago. The synthesisers of the 1980s landed in millions of homes, yet hardly anyone could program them. The power was there; the control surface wasn't. Agentic AI is heading for the same trap, only with far higher stakes because these systems now act on the user's behalf.
The control problem nobody is naming
An AI agent that books, sends, buys, deletes or edits things introduces a new category of risk: consequence without consent. Traditional software waits for a click. An agent runs ahead, chaining actions while the user watches — or worse, while they look away. When something goes wrong, the user's first instinct is primal: stop it now. If your interface can't honour that instinct instantly and legibly, you haven't built an assistant. You've built an anxiety machine.
The hard part isn't the technology. Pausing a process is trivial. The hard part is UX: making the stop affordance visible, trustworthy, and reversible so that everyone — not just power users — feels in command. Control is not a safety feature bolted on at the end. It is the product.
Four design principles for agents everyone can control
1. Make the stop button loud, not hidden
A stop control should never live two menus deep. It belongs in the primary field of view, available at every stage of the agent's run. Treat it the way a lift treats its emergency button: oversized, unambiguous, and impossible to confuse with anything else. If users have to hunt for it mid-task, trust collapses before the agent even makes a mistake.
2. Show the work before the consequence
Agents should preview their intended actions and pause at points of real consequence — spending money, contacting people, deleting data. This mirrors a quieter UX truth: everyone's got a quick question, and most of the time they want to confirm something before committing. A confirmation step at the right moment isn't friction; it's the difference between a tool and a liability.
3. Make every action reversible — or clearly flag when it isn't
Stop is only half the story. Users need undo. When an action can't be undone, the interface must say so plainly before execution, not after. Distinguish reversible steps (drafting, filtering, suggesting) from irreversible ones (sending, purchasing, publishing) with distinct visual language. Ambiguity here is where people get burned.
4. Design the controls for assistive technology from the start
As LLM-driven chat and agent interfaces spread, accessibility stops being optional. A stop button that a screen-reader user can't find, or a streaming response that never announces when it's finished, excludes people who most need reliable control. Accessible control surfaces benefit everyone: live-region announcements for status, keyboard-reachable stop and pause, clear focus order, and labels that describe what the agent is actually doing. If only sighted power users can steer the agent, you've rebuilt the unprogrammable synth.
Why this is a business issue, not a nicety
At Evocative, we treat agent control as a conversion and retention lever, not a compliance checkbox. Users adopt automation in proportion to how safe they feel handing over control. The brands that win the agentic era won't be the ones with the most autonomous features — they'll be the ones users trust to run unattended because they know they can intervene at any moment.
There's a commercial pattern worth noting: capability without controllability produces impressive demos and poor adoption. The synth that everyone owned but few could program sold once. The instrument that invited mastery built a culture. Agents are the same. Give people legible control and they'll let your agent do more over time. Hide it, and they'll disable the feature after the first scare.
A practical checklist for your next agent
- Persistent stop: visible and reachable at every step, including via keyboard and screen reader.
- Previews at consequence points: pause before anything costs money, reaches a human, or destroys data.
- Reversibility map: label every action as undoable or permanent, and warn before permanent ones.
- Status transparency: announce what the agent is doing, when it's working, and when it has finished.
- Accessible by default: live regions, focus management, and descriptive labels tested with assistive tech.
Autonomy is not the goal. Earned trust is. The stop button is where that trust gets designed — and where most teams are still leaving a blank space on the canvas.
Sources
- The synth everyone owned and few could program: a lesson for AI design — UX Collective
- Everyone is building AI agents. Who’s designing the stop button? — UX Collective
- Making Immersive LLM Chat Experiences More Accessible — Smashing Magazine
- Everyone’s got a quick question — UX Collective
