When your app starts thinking about you while you sleep
Mark Sage - 8 min read - 15/09/2026
Last Saturday morning, my phone buzzed before I'd properly woken up. It was Kith — an app I've been testing, still in beta — wishing me a happy Saturday and filling me in on things coming up in the week ahead.
It wasn't a calendar alert. It wasn't a reminder I'd set. It was a reach-out, from the machine, with personality and a little sass. And sitting there in bed, half-asleep, I realised I'd never had that experience with software before. Not a notification about something I'd scheduled — a notification that existed because the app had been doing work while I wasn't looking, and had decided, on its own initiative, that now was a good time to tell me about it.
That moment is what this piece is about, because I think it's the first crack of something much bigger.
Native, not ported
Kith is a business development agent. Its domain is leads, connections, marketing, follow-ups. What's interesting is that its entire interface is an LLM chat window. There's no dashboard, no real menus, no depth of functions and services. You simply talk to it.
That sounds like a small thing. It isn't. Because most "AI-powered" software shipping right now is the equivalent of a DOS app running in a Windows window. When DOS applications were first ported to Windows, they technically worked — but they were still DOS apps. Keyboard-driven, character-mode thinking squeezed into a frame. They weren't of the new platform; they were the old platform wearing a new costume.
Look around at most AI features today and it's the same story. A chatbot bolted onto a CRM. A sidebar assistant in your email client. Underneath, nothing has changed. It's still a fixed database schema, some screens, and a conversational veneer. The UI remains the system of record, and the chat is a query language bolted onto its front.
Kith feels different, and I think the difference is architectural rather than cosmetic.
This isn't software built a decade ago and strapped into a chat view. It was built from the ground up to understand chat — to take messy, ambiguous human utterance as its native input, extract the intent, and map it into a flexible structure built around the problem domain. Telling it about a conversation you had at a conference doesn't just log a note; it extends the model.
The chat isn't a way to query the database. The chat is how the database comes to exist.
A different way of working
Spend an hour with it and the practical strangeness settles in. There is no menu.
Every interface I've used in my working life has been a negotiation between what I wanted and what the software had pre-imagined. A traditional application exposes its structure — screens, tabs, forms, buttons — and I map my intention onto it.
Kith inverts that. I express intent in my own words, and the app maps itself onto me. Want to load in your LinkedIn contacts? Just say so — it responds with how, you drop in the file, and it handles the rest. Take a photo of a business card. Paste a screenshot. Speak a note after a meeting. It doesn't care about your format, only your meaning.
Every prior generation of software gated content by type: this screen accepts a CSV, this field accepts text, this button opens the camera. For the first time, format is the app's problem, not mine. The file is no longer something I must prepare for the software; it's something the software must make sense of.
What's enabling all this is the thing you never directly see. Behind the chat, Kith is building a map — not just of your contacts, but of the reasons they're connected. An event. A project. A past job. People are nodes, but so are the contexts that link them, and the edges carry meaning. It's a graph of your professional life that accumulates every time you interact with it.
And it combines that graph with your diary. It knows what's coming up. It knows you met someone three weeks ago who is relevant to Thursday's meeting.
This is what gives it something ChatGPT, in its vanilla form, simply doesn't have: prescience.
ChatGPT is like a porch light — it only shines when you interact with it, and it stops the moment you leave. Kith is closer to a colleague who keeps working between your conversations, notices things, and taps you on the shoulder.
Which is how I ended up getting wished a happy Saturday by a piece of software.
Here's the part I didn't expect, and the part I find genuinely disconcerting.
Kith doesn't feel human in the sense that you mistake it for a person. It feels human in the sense that, within a few exchanges, you start treating it like one. Specifically, like a colleague on a Teams chat, albeit one specialised entirely around business development.
The behavioural shift is automatic and fast. You start giving it less context than you'd give a tool, because you start assuming shared context — the way you'd say to a coworker "follow up with Sarah about the thing from last week."
You leverage the assumed memory, the shorthand, and you start directing your effort differently. With a tool, most of your energy goes into operating it; with a colleague, most of it goes into reviewing their work. That inversion — from operator to reviewer — is probably the defining economic shift of this whole category.
The sass helps. Personality is usually either cosmetic (a mascot) or grating (a paperclip). Here it serves a purpose because it's the surface signal that there's an agent with a consistent disposition on the other end, not a search box. And that's why you change how you interact - you actually say thank you, you ask what it thinks, and you feel a flicker of guilt when ignoring it. None of that makes sense for software. All of it makes sense when it moves from feeling like software to feeling like a colleague.
This is just the beginning
If all of this were just my experience with an obscure beta app, it would be a curiosity. It isn't.
On 26 August, Salesforce — the company that spent twenty-five years teaching the world to click through CRM tabs — announced Claudeforce, a partnership with Anthropic that puts Salesforce inside Claude. The first iteration gives sellers 37 pre-built skills — meeting prep, deal health review, pipeline review — running against live CRM data, through conversation, without opening Salesforce itself.
It's in pilot now, with open beta expected this month. Benioff's own framing was "the UI is the AI." His President of Applications went further, saying plainly that when people stop using Salesforce through the traditional interface and start using it through an agentic one, it increases the value of Salesforce.
Sit with that for a second. One of the largest SaaS vendors on earth just said publicly, that its customers may stop looking at its interface — and that this is good for business. The screens became optional. The system of record became the moat.
Salesforce's founding pitch was "no software," meaning nothing to install. In this new world it's more likely to become "no software to look at".
To be fair to Claudeforce, it's arguably still the DOS-app-in-Windows stage of this transition. Claude reasoning skillfully over twenty years of pre-existing tabs and objects. Valuable, but the ontology underneath is the ported one.
Kith is what the native version looks like when the domain model is designed for inference rather than retrofitted around it. The industry is currently shipping the port, but the beta on my phone is the native.
Apple sees it too
Look up a level, and the platforms see the same thing coming.
At WWDC in June, Apple's developer announcements were arguably more revealing than its recent consumer ones. The big theme was making apps' capabilities accessible to system-level intelligence: an expanded App Intents framework connecting apps to Siri AI, with personal context understanding, app actions and onscreen awareness, plus Foundation Models and Core AI frameworks that let developers plug their apps into on-device and cloud models — Apple's, or anyone's.
Or to put that more plainly, Apple is building the rails that let an intelligent layer act through apps, rather than requiring you to navigate them.
That's the same bet as Kith and Claudeforce, made at operating-system scale. The app exposes its capabilities to an agent; the agent, with its model of you, decides when and how to use them. You essentially stop being the router.
I found it telling that the recent product event, for all its hardware brilliance, had so little to say on any of this. The interface story is happening elsewhere in Apple's stack right now — in the frameworks, not on the stage.
But that touch-based, icon-grid, app-launching mobile interface we've all grown up with is nearly twenty years old. It has been the constant of modern computing for most of its existence. If the pattern of this piece holds — interfaces change when the underlying capability changes — it is due. Not abolished; demoted. Still there when you want to browse, but no longer the only way in, and no longer the most efficient way to get the job done.
Where this leads
Strip it down and three things are true at once.
The interface layer is dissolving. Menus were a brilliant compromise — a finite, visible map of everything software could do — but they also made everything software couldn't do invisible. Intent-based interaction removes the map. The app is no longer a place you go; it's a capability you invoke.
The structure is moving behind the scenes. An app that only exists while you're typing can never tap you on the shoulder — but as the graphs and context models that used to be the interface become plumbing, the app starts to remember. And memory, persisted, becomes proactive. That's what sent the Saturday-morning notification.
And the relationship changes. Software you operate becomes software you collaborate with — specialised, tireless, slightly cheeky colleagues, one per domain, each with its own model of its corner of your life. That will be extraordinarily useful. It will also be quietly disorienting, in ways we'll spend the next decade noticing.
Saturday morning told me so.