Newly

Newly’s agent writes, runs and ships mobile apps from natural language. That is a lot of machinery to put behind a text box: sessions that run locally or in the cloud, a simulator, git branches, model choices, deployment to TestFlight and the App Store. The design problem was deciding what a person needs to see, and when — and keeping the rest one click away.

I joined as the designer on a small team. The constraints were real: desktop and web had to feel like one product, the audience included people who had never opened Xcode, and the agent’s work had to be visible without turning the chat into a terminal.

My first direction treated the app like a product roadmap. The agent proposed features as cards — “Proposed by Newly”, with a reasoning panel citing session data and drop-off rates — and you approved, built and deployed them from a board. I pushed it far: a git-style tree of sub-features, a “New Idea” card you could compose inline, and three columns for Features, Marketing and Website.

The workspace is a sessions sidebar, a conversation and a device preview — each a rounded card that can collapse on its own. Pinned and recent sessions sit on the left with a cloud or laptop glyph for where they run and a blue dot for the one that is running; the conversation stays in the middle; the simulator takes the right.

The prompt bar is the one element on every screen, so it got the most attention. It has three modes — Build, Plan and Ask — in a segmented control where only the active mode shows its label, so the bar stays narrow but always says what it will do. Attach on the left, microphone or send on the right, and a second row for the things you set once: where the session runs, which branch, which model.

The conversation is deliberately plain: your messages in bubbles, the agent’s as text, and its work folded behind one line — “Worked for 3m, used 3 tools” — that expands to the commands it ran. The device pane starts as an explanation and a Run button, then becomes a live iPhone 17 Pro with a rail of simulator controls beside it.

Deploy groups destinations by platform — iOS, Android, Web — with the public release in blue and testing targets in grey. Pick the App Store and a release card slides in with a short checklist and, underneath, what the agent already filled in. The point was to make the agent’s busywork legible, not invisible.

The desktop app opens like an IDE: a sign-in splash, then a small welcome window with recent projects. On the web the same workspace runs in the browser, with a brand gradient and a “Pick up where you left off” toast. The two differ only where the platform does — traffic lights here, a download nudge there.

The interface is nearly colourless on purpose: an off-white window, white cards with hairlines, one typeface, Phosphor icons. Blue is reserved for the active mode, the primary action and anything that is running. The brand’s pink and the gradient wallpapers stay in the logo, the launcher and marketing, so the user’s own app is the most colourful thing on screen.

Most of the work was states, not screens: what the bar says while the agent connects, fixes a bug, or holds your next three messages. Designing those honestly — a status line, a queue, a stop button — did more for trust than any onboarding could.

I also learned to kill a direction I liked. The feature board was the most “designed” thing in the file, and the right call was to keep three ideas from it and let the rest go.