Patterns

A two-step wizard card for naming an agent, picking its icon, and selecting which event sources activate it.

A configuration panel for listing and toggling the external events that activate an agent, with a manual run control.

A compact dialog that pauses an agent mid-task to collect a structured choice — radio options, a free-text fallback, skip/continue controls, and multi-step pagination.

A side drawer that tracks an agent's background work live — running tasks with elapsed time, stats, and stop controls, plus an expandable history of finished commands and output.

A numbered instruction editor where typing / opens a two-step inline picker — choose a tool, then its connector or target — inserting an atomic, editable chip into the prose.

A doc-style editor for an agent's step-by-step procedure — a / step picker, IF / ELSE branches with plain-language or code conditions, section jumps, and escalate / switch actions.

A live, agent-authored plan — an ordered list of steps marked pending, in progress, or completed as the agent works through them.

A compact run card with the agent's current step and a timer, a segmented progress bar, and a list of finished steps that stacks up as it reads, edits, tests and pushes.

A collapsible settings card for what an agent does when the customer goes quiet — a wait duration, a Team / Workflow handoff target, the handoff message, and a working-hours option.

An editable plan shown before an agent runs — reword, reorder, remove or add steps, or ask for a new plan — that locks once approved.

A tree of parallel sub-agent tasks, each with its own status and step list.

A collapsible "Thought for Xs" summary that expands into a step-by-step trace.

A live, timed header over reasoning that streams in as a step timeline, then folds into a "Thought for Ns" trace when the answer starts.

A button with an animated light sweep across its surface.

A send button that morphs into a stop control mid-stream and back on completion.

A loader with a shimmering label and a live elapsed-time counter.

A line icon that draws itself stroke by stroke while the step it stands for is running.

An inline pill naming an in-flight tool call that resolves into a result summary.

A full-width indeterminate progress bar over a list whose items resolve from skeletons one by one.

A streamed answer with inline sources, actions, and follow-ups.

A hoverable, clickable footnote-style source marker inline within text.

An overlapping stack of source favicons with a count, expanding into a linked source list.

A paginated single-source card with a trust badge and domain pills for jumping between sources.

A vertical list of suggested next questions shown after a response, each sendable with a tap.

A floating toolbar on text selection for describing edits, explaining, or improving a passage.

One-tap adjustments (Rephrase, Shorten, More formal, Bulletize…) and a free-form instruction field docked under an AI output, with every result kept as a steppable version.

A composer with @ sources, / commands, a model picker, and dictation.

A first-run composer with suggestion chips, @ sources, / commands, an environment picker, and an orchestrator picker.

A hidden bottom-edge handle that wakes into an action dock on hover and summons a floating prompt bar.

A composer that keeps working mid-turn: ⌘Enter queues follow-ups above the input to send now, edit, or auto-send when the agent finishes.

A composer chip that opens a list of models, each with a one-line description, plan-locked options that lead to an upgrade, and an extended-thinking switch.

A header control that starts a temporary chat, with a persistent banner stating exactly what won't be saved, remembered or used for training.

An Enhance button that rewrites a short prompt into a specific one in place, highlighting what it added, with one-click undo.

A prompt written as a sentence with inline blanks — free-text slots that grow as you type and choice slots — that runs once the blanks are filled.

A collapsed summary of a batch of file edits, with undo and an overflow list.

Per-file chips that switch an inline diff viewer between a batch of changed files.

A code block that flips between the source and a sandboxed live rendering of it.

An auto-scrolling, collapsible panel streaming raw command output line by line.

Workflow trigger and condition steps on a dotted canvas.

A message bubble with feedback thumbs, edit-and-resubmit, retry/regenerate, and a version stepper.

A short card of questions before an ambiguous task, each pre-filled with the assistant's suggested answer, that collapses into a summary once started.

Editing a sent message creates a new branch instead of overwriting it, with a ‹ 2 / 3 › switcher that swaps the message and its reply together.

A pending tool-call prompt with allow, always-allow, and deny actions.

A settings panel for connecting an app or MCP server — shows capabilities overview, groups tools by type, and lets users set per-tool permission levels.

A panel listing what the assistant remembers about the user — edit or forget each memory with undo, switch memory off, forget all — plus a memory-updated notice for the chat.

A composer's file/image attachment tray with drag-drop, upload progress, and inline preview.

A row of tiny image thumbnails that grows into an anchored preview card with a label, counter, and prev/next stepping.

A quota-exceeded state with a live countdown to when the user can send again.

A cut-short assistant reply with Continue and Retry actions.

A console-style error card for a failed generation, with details and retry.

A structured card surfacing a browser console error with a code frame, call stack, and navigation.

A dismissible card surfacing a network connection failure with a description and retry/details actions.

A side-by-side pair of response panels with independent regenerate and a single preferred pick.

A badge or inline dot signaling how sure an AI answer or extraction is.

An AI-generated label in chip, image-overlay and inline forms that opens a provenance card: model, time, and whether a person edited it.

A pre-run estimate of what a task will cost, with a live balance meter, a low-balance warning, a top-up when the balance can't cover it and a breakdown.

A context-window meter: tokens used against the limit, a segmented bar of what fills it, a legend to inspect each source and a Compact action to free space.

A live amplitude waveform while the AI listens or speaks.

A 3D particle sphere that breathes and spins with the voice conversation.

A pulsing mic indicator showing the AI is actively listening.

A streaming, speaker-labeled transcription of a voice conversation.

A searchable ⌘K overlay for jumping to sessions or running quick actions.

A stepped onboarding card with circular arc progress, dashed-circle pending states, and strikethrough for completed steps.

An avatar stack showing who has access to a shared record and who is actively viewing it right now.

A two-step modal for searching or emailing people to a record and assigning each a role, with a live current-members list.

A setup screen that hands users a pre-filled prompt to paste into their AI coding agent, with a snippet fallback and a live listener that confirms the first event.

A floating pill that copies a setup prompt for the user's AI coding agent in one click — outcome label, a row of supported agents, a tooltip that says what the click does, and a separate dismiss button.

A Describe button in a table toolbar that turns a plain-language request into removable filter chips — column filters where a column exists, marked semantic matches where none does.

A record drawer of key/value fields where the model fills the empty values — each arrives as a sourced suggestion to accept, edit or reject.

A SQL-first query bar that also takes plain language — the mode badge flips to Ask, and Enter translates the request into a visible, editable WHERE clause before running it.

A side panel for checking extracted values one by one — each field shows its value, the model's reasoning and verbatim evidence that highlights in the source, with accept, dismiss and apply-all-pending.