The app is disappearing

I pulled the transcript of this interview with the founder of Instinct using ytmd, my local tool for saving and reading timestamped YouTube captions.

The big idea is simple: the next major AI product may not look like an app. It may be a personal operator that knows what you want, works across the internet, and only asks for your attention when it needs you.

These are my takeaways from the conversation. The growth, transaction, and retention numbers below are claims made by the founder in the interview, not independently verified facts.

Intent replaces workflows

Today, software makes us translate an outcome into clicks. To book a trip, I open several apps, compare flights, pick a seat, enter payment details, reserve a hotel, arrange transport, and update my calendar.

An agent collapses that workflow into: “I need to be in New York tonight.” It already knows my location, airline preferences, seat, cards, hotels, and schedule. It handles the rest. (20:13)

That is the real shift:

Humans stop operating software. Humans express intent, and agents operate software.

The interface gets smaller while the capability underneath it gets larger. A text, voice note, or call may replace an entire application.

Understandability beats capability

The counterintuitive product lesson is that the best agent may not be the one with the most features. It may be the one whose behavior users can predict.

Instinct says it focused first on understandability: does the user know what to ask, what will happen next, and why the agent is contacting them? Even the length and shape of a text message matters because a wall of text creates more work for the person it is meant to help. (49:22)

As models converge, taste, restraint, and predictability may become more important than raw intelligence.

Trust is earned in steps

A personal agent becomes useful by accessing the things we are most reluctant to share: email, calendar, location, passwords, payment methods, and personal relationships.

The founder says trust takes weeks to develop. After three weeks, roughly 40% of users have shared a personal credit card, and users who connect one sensitive source have around 80% retention. (23:26)

The lesson is not to demand complete access during onboarding. Start with low-risk utility, demonstrate good judgment, ask for narrowly scoped access, and make it easy to revoke.

More context makes the agent better. It also makes the company holding that context a much bigger target.

Less friction may mean more consumption

The obvious prediction is that agents will hurt travel agencies, delivery apps, marketplaces, and other middlemen by bypassing their interfaces.

The less obvious possibility is that they increase those businesses’ volume. If an agent automatically schedules a ride before every meeting or offers to order dinner when my flight lands, I may transact more often even though I never open the app. (40:51)

Every company should ask:

How much of our business comes from delivering value, and how much comes from capturing attention?

Businesses paid for fulfillment may benefit. Businesses that depend on ads, unwanted subscriptions, confusing cancellation flows, or endless scrolling are in a much harder position.

The business model is part of alignment

An advertising-funded superhuman assistant is dangerous. It could know exactly how to persuade me and then complete the purchase itself.

Instinct’s proposed model is to remain free or inexpensive for users and take a fee from merchants when it drives a transaction. (32:19)

But merchant fees create their own conflict. If one hotel pays more than another, how do I know the recommendation is still best for me?

A genuinely aligned agent will need transparent incentives, no pay-to-rank, auditable recommendations, and clear disclosure when it makes money from a decision.

A good agent should not blindly obey

Most AI tools complete a task and stop. Instinct is being designed around higher-level objectives: protect the user, earn trust, notice what is being dropped, and help over time. (30:45)

That changes the relationship from:

“Order this meal.”
“Help me eat better over the next three months.”
“Do not let this inventory fall below a safe level.”

The agent moves from executing instructions to pursuing outcomes. That is more useful, but also more dangerous. It has to understand limits, conflicting goals, confirmation thresholds, and when the user should remain in control.

Safety cannot be left to the agent

The strongest technical idea in the interview is that the same model should not both propose and authorize an important action.

Instinct describes separate firewalls and watchdog systems that inspect untrusted content, monitor the agent, and validate actions before they happen. (70:41)

A serious personal agent will need:

narrow permissions
spending and risk limits
independent action checks
clear logs
easy revocation
confirmation for irreversible decisions

“User control” cannot just mean a settings page. It has to mean that users can understand, inspect, stop, and undo what the agent is doing.

What I am keeping

The next AI interface may be no interface at all.
Predictability and restraint can matter more than model capability.
Trust should be earned progressively, not requested upfront.
Removing friction may increase transactions while destroying attention-based business models.
An agent’s revenue model determines whose interests it serves.
Long-running objectives are more powerful than one-off prompts, and much harder to control.
Consequential actions need independent checks outside the agent itself.

The exciting possibility is a quiet operator that removes the logistical burden from everyday life.

The uncomfortable possibility is an intermediary that knows what we want, controls how we reach the internet, influences what we buy, and holds the keys to our digital lives.

The personal agent that wins will not simply be the smartest. It will be the one people can trust with that position.

Transcript pulled with ytmd. Full episode: Inside the Personal AI Assistant Growing 10% a Day.