A theory I'll defend at any whiteboard: the agentic AI race will not be won by the most capable agent. It will be won by the most delegable one — and delegability is a user experience property, not a benchmark score.
Think about the best human delegation you've ever experienced. A great colleague taking work off your plate doesn't just do the task. They tell you what they're about to do, check in at the moments that matter, flag what they're unsure about, and hand back something you can inspect. The skill isn't only competence; it's legibility. Agents need the same skill, and right now most of them are brilliant interns who vanish for an hour and return with a deployed mystery.
The heuristics never left
Here's the optimistic part: we are not starting from zero. The usability heuristics our field has refined for three decades map almost one-to-one onto agent design. They just need translation.
- Visibility of system status becomes visibility of intent. Before an agent acts, I should be able to see its plan at a glance — not a wall of reasoning, a plan. During the work, a heartbeat: what step, what's left, what changed.
- User control and freedom becomes undo for actions, not just text. The single biggest unlock for trust is reversibility. An agent that works in drafts, branches and staging areas can be given real responsibility. An agent that acts irreversibly must crawl.
- Recognition over recall becomes receipts. After the work: a diff, not an essay. Show me what existed before, what exists now, and which decisions were judgment calls.
- Error prevention becomes asking at the right altitude. The worst agents ask nothing; the second-worst ask everything. The design problem is calibrating which decisions are theirs and which are escalations — exactly the calibration a good manager does with a new hire.
The trust gradient
The mental model I keep coming back to is a gradient, not a switch. Nobody hands a new colleague the production keys on day one — trust is climbed in increments: watch me → do it with me → do it and show me → do it and tell me → just handle it.
Autonomy should be earned in the interface, the way it's earned in a team.
Today most agent products offer two settings: a chatbot on a leash, or full autonomy and a prayer. The products that win will let users move each category of task up and down that gradient independently. I might trust an agent to triage my inbox unsupervised, draft replies with review, and never send anything on its own. That's not a feature matrix — that's an employment contract, and someone needs to design its interface.
What this means for designers
If you design products, agentic AI is the biggest expansion of our remit since mobile. Concretely, new artifacts need designing:
- Plan previews — the moment between intent and action is the new most important screen in software.
- Progress narration — status for a process that thinks. Neither a spinner nor a transcript; something honest in between.
- Review surfaces — diffs for non-code work: "here's your calendar before and after", "here are the three emails I'd send."
- Permission ergonomics — granting capability should feel like lending a key, not signing a EULA.
None of this is exotic. It's information architecture, progressive disclosure, and honest feedback — our oldest tools, pointed at our newest material.
Why I'm optimistic
The pessimistic read says agents make interfaces irrelevant. My read is the opposite: agents make interface quality decisive. When capability is abundant and roughly equal — and it increasingly is — the differentiator is whether a human can confidently hand work over. Confidence is built by design.
Engineering made agents possible. Design will make them trusted. And trusted, not possible, is what people pay for.