Is the interface dying?
For thirty years, the product was the place. Today, the product is one of many 'places' you connect with your customer. What does this mean for the interface?
Two weeks ago I set up Instinct and started messaging with it directly in iMessage.
This is a product (an agent) without an interface. It's showing up in another surface, and taking full advantage of the features and feel of the iMessage (or WhatsApp) ecosystem - almost like you wish your friends would:
Read receipts on
Near-instant response with reacts or 'typing' indicator
Strong 'react' game. From a list emoji 📋 to signal "added to your to do list", or a check emoji ✅ to signal "it's done"
It's not overly verbose (although sometimes the 4 back to back messages can be a bit much)
Messages are sometimes sent with effects (nice touch)
My gmail is connected, it's helping me with life admin and personal to do's. It's clear that the once fictional idea of 'personal agents' is here. So what does this mean? For your product. For design. For interfaces...
Are interfaces still important?
For thirty years, the product was the place.
An app, a dashboard, a set of screens - usually on top of one or many databases.
Designing the product meant designing that place, and an entire discipline called UX grew up on the assumption that the screen was the product and everything behind it was plumbing.
Three things are slowly eroding that assumption...
No interface. Instinct, the agent I tried, has no app. You text it or call it in iMessage, and on September 8 it started giving every user's agent its own email address so it can open and run accounts on their behalf. A product with no interface (or let's just say a 'borrowed interface') that becomes a user of every other product.
Outsourced interface. Salesforce in Claude lets a seller review pipeline, prep a meeting, and update records without opening Salesforce. The interface became conditional: Claude when you're thinking, Slack when you're coordinating, the full Salesforce app when you need to touch a record or collab with the team.
Voice interface. Voice has been a demo for a decade. OpenAI put GPT-Live-1, the full-duplex model behind ChatGPT's voice mode, into the API: it listens and speaks at the same time, handles interruptions, and hands the heavier reasoning and actions to a backend model. Google shipped a run of Gemini Live updates this month, including a guided-vision mode that talks you through what the camera sees. With voice, every design decision goes into behavior: when to speak, when to wait, how to interrupt, what permissions and action feel like.
The interface is no longer the only star of the show - from the fixed container of the product, to one of several ways an intelligent system can choose to show up.
Screens still matter, but the 'product' can now be experienced somewhere else and in other ways - and those need a design practice.
In the book we called it AX, agentic experience, an extension of UX for the AI age.
A basic duality in 2026
On March 30, 2000, Jesse James Garrett published a single page called The Elements of User Experience. Some may remember the stack: user needs at the bottom, visual design at the top, with the flows and structure and controls in between.
Fewer people remember the headline at the top of the page: a basic duality.
He wrote that the web had been conceived as an information space and had turned into a software interface, which was a little confusing to everyone because, while they're different, everyone just ended up mushing everything together under 'the web'.
The diagram linked above exists to hold the two natures side by side.
We're now facing a new, similar duality because of agents.
The product was conceived as a place a person goes, and AI has turned it into something that acts on a person's behalf. Two natures again. And the confusion is the same kind: designers trying to describe behavior through the vocabulary of screens.
So here is the page, redrawn in the same style, but for 2026.
The six 'new' layers of design, through one journey
Rather than define the layers in the abstract, here they are in one situation, bottom to top, with what design produces at each.
A business traveler is in a taxi at 4:30pm.
Her 6:15pm flight home just got cancelled.
She has a board dinner tomorrow at 7.How does her agentic travel product show up?
1. Understanding: intent and context
Getting on another flight is the surface request. Being home for tomorrow's dinner, rested, without an evening on hold, is the real intent.
The system also needs to know what it's allowed to know: her calendar, loyalty status, travel policy, seat preferences, the weather at the connection.
Salesforce in Claude does a version of this before it shows a seller anything: it reads their CRM, Slack, and connectors during onboarding and builds the first view from what it finds.
The system learns the person before the person learns the system.
Design produces: an intent map (the surface request, the real job, the constraint) and a context inventory. The inventory might be a table with three columns.
2. Judgment: prioritize, refuse, act or wait
This is a new, important layer. Judgment used to be ours alone (yes us - humans!), decided once and baked into flows.
But now it means:
prioritize arrival time > price within policy
avoid trying to rebook the dinner, and
refuse the 5am flight, that technically works, because her preferences say never
It also means deciding whether to act now or wait.
This is the layer Jev from Typesafe is built for. State in, a decision and a calibrated confidence, no convo needed.
Whatever you make of the launch claims, a model that only makes decisions based on probability and defined set of options is a strong hint that judgment has become its own layer, with its own economics.
Design co-produces: a behavior spec and the evals that hold it to account. Evals are the new acceptance criteria.
3. Delegation: what it may do, and when it asks
She has said: "rebook me within policy without asking, but always ask before spending my money or changing a hotel."
Her company's policy adds a second source of authority, and maybe anything legal is the third. Those guidelines and hierarchy are critical for trust, and for designing an experience that feels natural and delightful.
Instinct's first week on the market showed what happens when authority is implied rather than designed: a tester publicly disconnected her email after the agent sent a message without explicit permission. Tsk tsk - no trust built, and being too eager led to missing the mark on delegation.
Design co-produces: the 'envelope' of guidelines as a policy, readable by the person and enforced by the system, plus the screen where she edits it.
4. Orchestration: tools, agents, humans, handoffs
The airline's API, the hotel's, the expense system, the corporate travel platform, and if the airline's system is down, a human who picks up the thread with everything above instead of asking her to start over.
Under the hood, it's a mix of touchpoints: deterministic code, a model deciding, two agents doing their own multi-step work, and a person on standby. The design question sits above it.
Does she meet...
One counterpart with the rest invisible, like Instinct?
One counterpart with the team showing, like Codex's "3 agents working" pill?
Many agents, like Grok Bot, where she picks who to talk to?
This could be the same plumbing, but it's three different relationships. This is where that choice becomes an interface. And it can't be made from the top alone. Design has to understand how the system works, then ask what the customer would do with the detail.
Design co-produces: a map of what's underneath and who she meets, and a decision, per product, about who she can talk to. The human sits at the top of that map as the premium tier: faster agents for most of the trip, a well-trained person for the messy and the emotional.
5. Trust and control: explain, approve, undo, recover
As the agent is doing it's thing, showing the 'chain of 'action'' can bring transparency (and piece of mind) to the experience. Transparency builds trust.
"Searching for Premium Economy" "Applying your loyalty number" "Prioritizing Delta Airlines"
One of the principles from our book on AX is 'transparency, tapered over time', but as the agent builds trust with you, like a great new employee, the visibility into what's happening behind the scenes can fade away.
Showing 'why' something was done can also help bring piece of mind to the recommendation. This isn't a new pattern or behavior, but now that an agent is acting for you, this design layer helps users proceed with confidence, or pause and look for alternatives.
Here's Serko AI searching for a trip to San Francisco. While the itinerary loads, it shows its work: applying my airline preferences, fewer stops, cabin class, my Aeroplan membership, then on the hotel side - boutique, in-room Wi-Fi, a gym, my Accor membership.
By the time the results land I already know what they were built from, and what to argue with if they're wrong. That's the difference between a spinner and a receipt.
Design produces: a clear definition of transparency needed for trust - both during the action, and for the results themselves. This can change with user maturity, and as they work their way along the trust curve.
6. Expression: surfaces for people, and for agents
The expression of the agent shows up across several of the layers we spoke about, but this is also where we need to decide what the core surfaces will be, and how your product will show up for the user at different stages in their journey.
Will you package up some of your functionality into an agent that can meet users in iMessage or WhatsApp... maybe Slack or Teams?
Will you pass through a ton of data via MCP so a ping to Claude can access part, or all, of your world?
Will you have an app where you leverage generative UI for select queries, dynamic UI for different users, or more traditional UI/UX with consistent design patterns tocreate a delightful experience 'at home'?
The reality is consumer behavior is changing, and along with it, their expectations are evolving.
The modalities are now all viable at once: message, voice, conversational UI, generative UI, traditional UI, borrowed UI... The condition selects the surface.
And there's a second audience: agents need a surface too, in the form of APIs, MCP, and structured offers, because a product an agent can't reach will get skipped no matter how good its screens are.
Don't forget: memory and learning
Memory changes these layers over time. Remember she prefers morning flights and never routes through Chicago in winter. Forget the card she used once. Make each cancellation smoother than the last. The same six layers on trip one as trip ten, different settings and feeling as you feel seen, heard, and well taken care of (digital hospitality).
Design co-produces: a memory schema with a never-store list, and the ledger the person can read and edit. Forgetting on request is as much a feature as remembering.
Where this leaves the interface...
It still matters, a lot. More surfaces. More understanding. More focus on trust.
When the interface was mandatory, mediocre screens survived because people had nowhere else to go.
But now, every screen competes with the option of no screen at all. The ones that appear have to be worth choosing.
Some products will be destinations on purpose, good enough that people choose to spend time there, and those deserve more craft than ever. The rest will show a screen at the highest-leverage moments in the system: the approval, the exception, the place a person builds a mental model of what's being done for them.
If you lead design or product
Three things worth doing:
Pick one real workflow and draw the 6 AX layers for it. Most teams discover they've never written down any of this, and have never decided how the product needs to show up, which moments deserve a screen, and more importantly how agents, if at all, will be incorporated into the experience.
Write five evals for the judgment layer before anyone writes a prompt. If you can't say what "good" looks like in a given situation, the model is deciding it for you.
Add one component to your design system that don't exist yet: explain a decision, ask permission, hand off to a human with context...
The interface isn't dying, but it's definitely evolving.
Don't let it evolve without you.