AI should be present without being summoned
The chat box makes you stop and decide what to ask. An ambient layer reads the room instead.
The chat box is the wrong mental model. Every time I open one I'm doing two things: stopping what I was doing, and deciding what to ask. That pause is where most of the value gets lost.
The AI agent literature makes a distinction worth taking seriously. An agent isn't a model you query; it's a software entity that perceives its environment, decides, and acts. Perception is the operative word. Something that perceives doesn't wait to be addressed – it reads the room.
Most of what ships today ignores this. The default pattern is a sidebar or modal: type a prompt, read a response, close the window, resume work. The model is brilliant; the interaction pattern is a secretary you have to schedule. You carry the overhead of knowing when to consult it and knowing how to ask. For most people in most workflows, that overhead is just enough friction to make the habit fail.
Ambient presence is the alternative. Not a persistent chatbot floating in the corner, but a layer that tracks context across what you're already doing – what tab you're on, what file is open, what you wrote ten minutes ago – and acts when it's useful rather than when prompted. The agent that notices you've been rewriting the same paragraph four times and offers a concrete alternative is more useful than any model you'd think to summon at that moment.
The toggle in this piece tries to make that concrete. On the left, a chat-box flow: you stop, switch context, compose a query, wait. On the right, an ambient trace: the agent reads what's in focus, surfaces something specific, stays out of the way when there's nothing to say. The two have the same underlying model. The difference is entirely in when and whether you're interrupted.
same document, same paragraph being rewritten. Left: a modal opens over the work and the cursor leaves. Right: the work stays in view and a marker appears in the margin.
The summoning pattern persists because it maps to how we used to think about software – you invoke a tool, the tool does a thing. Agents that perceive and act don't fit that frame. They require trusting that something is watching without being intrusive, which is a different kind of UX problem than building a good prompt interface.
The better versions of this are starting to appear. Not as products that call themselves agents while fronting a chat box, but as integrations that stay below the surface until they have something worth saying. Building the right ambient layer is harder than building another chat UI. It requires deciding what the agent should notice, what it shouldn't, and when its threshold for acting is low enough to help but high enough to avoid noise. That is actually the product.