Quests, not feeds

Search engines solved much of that. Then feeds arrived, and the problem flipped: there was too much, and the ranking was done by companies paid for your attention. Generative AI pushes this further. Text, images, video and messages can now be produced at almost no cost, in any volume, tuned to whoever is watching. A growing share of what reaches you will be written by machines, and some of it will be written by machines trying to persuade you.

In that world, the useful question is not "where can I find more?" but "what, out of all this, actually needs me?" A bill that is about to go up. A form with a deadline. A message from someone you care about that got buried. A decision that will be much harder next month than today. These are easy to miss, not because they are hidden, but because they are surrounded.

People already rely on filters to cope: friends, editors, assistants, habits. The trouble with today's main filters is whom they serve. A feed ranks by what keeps you looking, because that is what pays. We want a filter whose only loyalty is to the person using it.

The player sees a small window, a "system" panel, that lists what is happening and what they might do next: a handful of quests, each with a reason and a next step. Nothing scrolls forever. The window is there when you want it and quiet when you don't.

We think a personal interface could work the same way. Your agent, which runs for you and keeps your data on your own device, looks at what it is allowed to see: your accounts, your messages, your calendar, answers from other services and agents. It picks out the few items that matter and shows them as quests. Each quest says what it is, why it is shown, and what the smallest useful next step would be.

• Strictly opt-in, and revocable. The quest window is off until you turn it on. You choose which sources it may read, and you can withdraw any of them, or turn the whole thing off, at any time, without losing anything else your agent does for you.
• You can see why. Every quest carries its reason in plain words: "your phone bill rose 18% this month", "this deadline is in four days", "you asked to hear from this person". If the agent cannot give a reason, it should not show the item. Numbers, dates and matches come from ordinary, checkable code, not from a language model guessing.
• Few items, then quiet. The window shows a small set and then stops. Finishing it is the goal. There is no endless scroll and no "one more thing" to keep you there.
• No engagement tricks. No streaks, no badges for logging in, no counts of time spent, no notifications designed to pull you back. The agent is not measured on how long you look at it. If anything, it succeeds when you need it less.
• Tuned by feedback you choose to give. You can tell it "more like this", "not now" or "never this". You can also, if you want, tell it how something made you feel: that a reminder was a relief, or that a topic left you anxious. Nobody has to give feedback, and the window still works without it.
• Feelings never leave your device. Emotional signals are the most intimate data there is. Whatever you share about how you feel is used only by your own agent, on your own device. It is never sent to a remote model, never sold and never pooled.

The agent behind the window is the same personal agent described in our other notes: a bounded assistant that keeps your data with you and answers only to you. When it needs something from a service, it asks a specific question and gets a specific answer or proof back, rather than handing over your records. The quest window is simply the place where that work becomes visible to you.

It can say "this charge looks like a subscription you stopped using; here is how to cancel it", but it cannot cancel anything unless you have explicitly allowed that kind of action. It can surface a message you would want to see, but it does not decide what you believe about it.

That line matters most on questions of value and belief. Your agent can warn you when something looks like coercion or manipulation, for example a message that uses your religion, identity or history to pressure you into paying or acting. It does not judge, rank or try to change what you believe, and it does not nudge you toward anyone's preferred opinion, including ours.

This is research and design. The first step is modest: a personal assistant that runs on your own device and handles one narrow job, such as finding recurring charges in your own bank statements and explaining them. That is a natural first source of quests, because the reasons are concrete and the numbers can be checked.

From there, the questions we want to work on are practical ones. How few items are enough? How should reasons be phrased so people can actually check them? How can someone see and reset what the agent has learned about them? How do we test whether the window reduces overload without measuring people in ways that would themselves be intrusive?

If you work on calm technology, interface design, recommender ethics or on-device AI, and you see a flaw in this approach, we would like to hear about it.