18 components18 new, October 25, 2026
AI
Interfaces for AI products and agents: generation, streaming answers, reasoning, agent runs, plans and permissions, knowledge sources, evals, enrichment, cited replies and chat history.
Image Generation CardNew, Oct 15
AI image panel: Generate resolves the image from coarse pixel blocks to sharp with a step counter, then four variations land. Pick one, or change the ratio.
Agent QuestionsNew, Oct 21
The questions an agent asks before it starts: pick with a click or a number key, and answers fold into a summary you can edit.
Docs SearchNew, Oct 16
Search for a docs site that can also answer. Hits are grouped by page with the line that matched, and the first row asks instead: the answer is written from the matching sections, each sentence cites its source, and pointing at a citation lights up that source. With nothing to cite, it says so instead of guessing.
Run TraceNew, Oct 25
One agent run as a waterfall: model calls, tool calls and retrievals stream in on the run's own clock, so parallel calls sit side by side and a timeout stands out. Hover to read the clock, click a span for its input, output, tokens and cost, and collapse a step into one bar that still shows where its children ran.
Tool PermissionsNew, Oct 25
Allow, Ask or Block for every tool an agent can reach, grouped by connected server and tagged as reads, writes or destructive. Setting a whole server to Allow leaves its destructive tools as they were until you say so, and changed tools show what they were until Save.
Knowledge SourcesNew, Oct 25
The files an assistant answers from, each drawn as its chunks. Cells fill in as a file is embedded, a scanned PDF says it needs OCR, and a test question lights up the chunks it would retrieve with their scores and matched words. Add the missing source and the ranking moves to make room.
Prompt TemplateNew, Oct 25
A versioned prompt where every {{variable}} is tinted in place and gets its own field, added and removed as you type. Point at a field to see where it's used, switch to Preview for the prompt the model gets, and save the edit as the next version.
Plan ReviewNew, Oct 25
The plan an agent writes before it starts: drag or arrow-key steps into order, rename, skip or add them, or ask for changes and see what the agent added. Approve runs it step by step, and steps that ask first wait for Run or Skip.
Streaming TextNew, Oct 25
An answer as it streams in. Words arrive in small uneven bursts and come into focus from a soft blur instead of popping in, with a caret riding the end. Stop freezes it where it is, and Continue picks up from the next word.
ReasoningNew, Oct 25
While a model thinks, its thoughts rise through a short window that fades at the top, and the seconds roll over like a counter. When the answer arrives the window folds into "Thought for 9s", which opens back into the full trace.
Response FeedbackNew, Oct 25
The row under an answer: Copy, thumbs up, thumbs down and Retry. A thumbs down turns the row itself into a one-line "What went wrong?" field, with the thumb moving into it. Enter sends and the field folds back into the row.
Eval GridNew, Oct 25
A prompt change run against its test suite. Cases resolve out of order as workers pick them up, and each is read against the old prompt, so regressions and fixes stand out even when the score doesn't move. Pick a cell for both outputs and the grader's reason.
Enrich TableNew, Oct 25
A list of companies with one AI column. Fill answers it row by row, says when it found nothing, and every answer shows the sentence it was read from on hover.
Agent InboxNew, Oct 25
The background tasks a team's agents are working through, sorted by what they need from you. Questions are answered in the row, and tasks slide between Needs you, Running and Done as they move. Open one for its step log.
Reply CopilotNew, Oct 25
A suggested support reply that writes itself in. Sourced sentences end with their article's number, and switching tone only fades in the words that change.
Chat HistoryNew, Oct 25
The conversation sidebar of an AI chat app, grouped by day. A new chat waits for its first answer, then writes in the title the model gave it.
Steer ComposerNew, Oct 25
An agent thread whose composer steers the run while it works. A steer waits above the input until the next step, then moves into the thread where the agent read it, and the remaining steps change to match.
Double ComposerNew, Oct 25
The prompt box of a coding agent in two layers: the card holds the prompt, model, effort and access, and a tray tucked under it holds the machine and branch. Moving the work to the cloud turns the branch remote in place.