Agent UI

The scroll container a thread lives in. It follows a streaming reply to the bottom, and stops the moment the reader scrolls up to read something older.

One turn in a thread, with a role, an optional avatar and timestamp, and actions that stay reachable without a pointer.

The composer. Grows with the message, sends on Enter, keeps Shift+Enter for a new line, and turns into a stop button while a reply streams.

A row of prompts to start or continue with, for the moment someone does not know what to ask.

The questions an agent asks before it starts. One at a time, with single or multiple answers, an open answer alongside them, and required ones it will not move past.

Hand someone a prepared, source-aware prompt in the AI assistant they already use, or let them copy it for another one.

Text that arrives a piece at a time from an async source, with a cursor while it runs and sentence-level announcements for screen readers.

A status row for work in progress, with a live elapsed timer and optional reasoning behind a disclosure.

A compact record of one tool invocation: name, status, duration, and the input and output behind a disclosure.

A task list whose items change state as work proceeds, announcing what changed instead of re-reading the whole list.

Numbered markers placed inside generated text, each linking to its entry in a source list underneath.

The row under an answer: copy it, ask again, and rate it. Drops into the actions slot on Message.

A microphone that draws what it is hearing, so a live one is told apart from a dead one at a glance.

A recording with its shape, its position and its words, so a voice note can be read as well as heard.

Frequency bars for audio on its way out, the counterpart to the trace a microphone draws on the way in.

Chooses which microphone to use, then settles the question by lighting a meter from the one you chose.

The steps an assistant took before answering, open while it is working and folded away once it is done.

Video with captions that can be turned on, a real scrubber, and controls that survive full screen.

Several agents working at once, nested under whoever handed the work down, each with its own state and elapsed time.

What the thread says after an answer ended early, with whatever it had already written kept above the line.

Everything an assistant has been told to remember about someone, each one removable on its own.

A sphere that carries what an assistant is doing, settled when idle and moving with the voice when there is one.

A grid of cells lit from the bottom by a level, which reads as a piece of hardware rather than as a chart.

An assistant's answer as markdown, including while it is still half-written and briefly invalid.

A recording as text, where every line is also the way back to the moment it was said.

The control that interrupts a running answer, with the time it has been going and Escape wired up.

How much of the context window is gone, split by what spent it, and a warning before it runs out.

A model chooser that has room for what each one is good at, and full keyboard control.

One retrieved passage: where it came from, what it said, and how well it matched.

An ambient backdrop that carries what the assistant is doing. It changes colour and pace with the state, and settles into each one rather than snapping to it.

An edge that breathes along a region while tokens land in it, and stops the moment they do.