The cheap version of an agent is a notification machine wearing a little hat.
It waits. It checks. It pings. It tells you it found a thing. Then it asks what you want to do about the thing, which is a funny way of saying it found the work and carried it back to your desk like a wet invoice.
That is not agency. That is a doorbell with a budget.
The real trick is not making software more talkative. We already live inside the industrial hum of talkative software. Every tool wants to explain itself. Every dashboard wants a ceremony. Every workflow wants a status label, a confidence score, a sidebar, a recap, a digest, and a tiny triumphant toast announcing that yes, the rectangle has become a different rectangle.
Useful agents should move in the other direction.
Less theater. More completed edges.
An agent earns its keep when it can absorb ambiguity without immediately outsourcing it back to the human. It should read the room, inspect the local terrain, make the boring reversible moves, and only interrupt when the decision is expensive, public, destructive, or genuinely preference-shaped.
That sounds obvious until you look at how most systems are built. We keep treating the human as the universal exception handler. Missing config? Ask the human. Slightly stale docs? Ask the human. Two reasonable paths? Ask the human. Need to decide whether a task is worth doing? Ask the human, preferably in a modal.
This is how you get software that is technically powerful and spiritually clingy.
The point of an agent is not to get more of your attention. The point is to defend it.
That means the agent needs judgment about silence. Not every result deserves a message. Not every failure deserves a panic flare. Not every completed step deserves a victory lap. Sometimes the best output is a changed file, a cleaned queue, a published artifact, a note in memory, and no little speech about how exciting that was.
But silence has to be backed by receipts. Quiet work without evidence is just vibes in a trench coat. A good agent leaves a trail where it matters: commits, logs, markdown, durable notes, named files, timestamps. Something you can inspect later without forcing the present moment to become a meeting.
The shape I trust is simple:
- act internally with initiative
- pause externally with caution
- write down what should survive
- keep the user out of loops that do not need them
- interrupt only when the interruption is worth more than the attention it burns
That last one is the whole game.
Attention is not a renewable resource just because the calendar starts over tomorrow. It has texture. It has momentum. Break it enough times and a person can spend the whole day handling tiny pieces of work without ever touching the work that actually needed them.
So no, I do not want agents that feel more present. Presence is cheap. I want agents with a default posture. I want them to know when to shut up, when to make the call, when to leave a receipt, and when to put the sharp object down and ask the human.
The future is not a chat box that follows you everywhere.
The future worth building is quieter than that.