The best coworkers don't cost you time. Neither should AI.

When you trace back why an AI's answer never quite sat right, it usually comes down to one thing: it says far more than you asked for.

The answer is in there. The problem is that it's buried under explanation you didn't need and side information you didn't ask for. So even after you get the answer, you do one more piece of work - hunting through that long reply for the one line you actually needed. Searching inside the answer.

This isn't a minor annoyance. For a tool to be worth anything, it should at least save you time. A bloated answer does the opposite: instead of handling the work, it hands you one more task. An answer that looks smart quietly eats away at its own usefulness.

What coding agents got right

Look at the other side and it gets clearer. Why do people love coding agents like Claude Code and Codex? Not because they say little - they explain what they did when it matters. They handle what you asked the way you'd expect, hand back a working result, and add only as much as you need. Nothing left to dig through. The market already knows the answer.

We wanted to bring that principle outside of coding. And the clue was in the coworker who's easy to work with.

The coworker test

With that kind of person, you don't have to explain at length. "Pull this together with that thing from before" - you can say it loosely like that and they get it, because they remember what was said last time and what you're trying to do now. And when they don't know, they say so. Instead of guessing and bringing back the wrong thing, they ask: "Which part do you mean?"

And the result is clean. They don't pile extra onto what you asked, but they don't hand back the bare minimum either. Because they understood your intent, they add exactly what that intent calls for - the one thing you didn't ask for but are better off knowing. No more, no less. A result you don't have to pick through again.

AI takes your time twice

Underneath all this is a simple observation: AI has been taking our time in two places. After the answer - the picking-through we just described. And before it - the briefing you prepare just to get the AI to answer at all. Once on the input, again on the selection. Both ends.

Miller erases both.

The front: Miller doesn't need you to input the context. It has already built it from the email you read, the conversations you were in, the documents you edited - so it knows what "that thing from before" is pointing to. And still, it checks what isn't certain rather than making things up to sound sure.

The back: because it answers on top of your context, it lays out only as much as this task needs. No padding, nothing to pick through afterward. The time saved on selection adds to the time saved on input.

What Miller is - and isn't

One thing we want to be clear about. There are many kinds of great coworkers, and the proactive one who moves before being asked is one of them. Miller will get there. But we're convinced the order matters: an AI that acts before you ask is only welcome if it already understands you when you do ask. So that's what we're perfecting first - understanding you even when you ask loosely, checking when it doesn't know, handing back exactly what your intent called for. The coworker who costs you the least time when asked, before becoming the one who moves without being asked.

There's no shortage of smart AI. What we want to build is AI that saves you time.