Almost every AI product ships the same screen. An empty box, a blinking cursor, and some version of "Ask anything."
The box is honest. It can do almost anything, as long as you know what to ask. That condition turns out to be doing a lot of work.
The work a chat box leaves with you
To get value out of a chat assistant, three things have to go right, and all three are your job.
You have to remember to ask. The assistant does nothing between conversations. If you do not open it, it does not exist.
You have to know there is something to ask about. This is the one that matters. The email that changes your quarter is not the one you go looking for. It is the one you do not know has arrived. A chat box cannot help with a question you do not know to ask.
You have to ask well. The answer is shaped by the question. A vague prompt gets a vague summary. A precise prompt means you already did most of the thinking.
None of this is a flaw in the models. It is a flaw in the shape. A chat box is a tool for someone who already knows what they need. Most of a founder's day is spent finding out what they need.
Moving the noticing
A background agent flips the direction. Instead of waiting for your question, it reads as things arrive, decides whether they matter to you, and tells you when they do.
The work that moves is the noticing. You no longer scan the inbox to find the one message that needs you. Something else scans it, and you start from its conclusions.
That is what Waveloom's feed is. It reads the tools you connect in the background. Most of what it reads produces nothing. When something needs you, it becomes a card with a one-line reason and, where a reply makes sense, a draft. You open the app and the thinking has already started.
Talking first is a privilege
An assistant that speaks first can also interrupt. That is the risk of the whole idea, and it is why the feed has rules a chat box never needs.
It has to be quiet. If it surfaces everything it finds interesting, it is just a second inbox with better prose. Low-confidence items are dropped, and there are hard caps on cards and notifications. I wrote about why in Silence is the product.
It has to explain itself. Every card carries the reason it exists. If the reason is weak, you can see that, and your dismissal teaches it.
It has to ask before it acts. Reading and drafting happen on their own. Sending never does. Every reply waits for your approval, and that rule is enforced in code, not in the prompt.
A chat assistant earns trust one answer at a time. A background agent earns it one silence at a time.
Where chat still wins
I did not remove chat from Waveloom. I moved it out of the centre.
Chat is the right tool when you do know the question:
- Following up on a card. Open its thread and ask for more context, or ask it to compare with something else it has read.
- Asking across your tools. "Is there anything from the Munich thread I have not replied to?" Waveloom reads across the sources you connected and answers with links back to them. See questions across your tools.
- Preparing for something. Before a call, ask for recent decisions and open questions from the connected context. See meeting preparation.
Chat reads by default. If you ask it to do something in a connected app, it shows you the exact action and waits for your yes, the same as the feed.
The test
Here is how I think about the difference. Close the app for a day. Open it again.
A chat assistant shows you the same empty box. It has no idea what happened while you were away, because you did not tell it.
A feed shows you what happened, what needs you, and what it decided to leave alone. Or it shows you that nothing does, and tells you how much it read to reach that conclusion.
The second one is the assistant I wanted. It is the one I am building.