Approval Request with Blast Radius
The agent stops and asks: the effect stated before the verb, choices as radios in a fieldset so picking is separate from committing, no pre-focused Approve, and the decision left on screen afterwards.
Tool-call traces, approval gates, context panels and inline actions.
An agent interface has a problem a chatbox does not: the model is about to do something, and a person needs to see what, and sometimes say no. These are the surfaces that make that legible — the tool-call trace, the diff-and-approve gate, the retrieved-context panel that shows what was actually read, and the inline "rewrite this" actions that live inside an editor rather than in a sidebar.
The agent stops and asks: the effect stated before the verb, choices as radios in a fieldset so picking is separate from committing, no pre-focused Approve, and the decision left on screen afterwards.
Machine-proposed row edits accepted one at a time, with before and after in real del/ins elements, a tri-state select-all that actually sets indeterminate, and a running count of what will be written.
The RAG debugging surface: ranked passages with similarity as a meter, matched spans in real mark elements, and filter chips wired as a radiogroup so arrows move between them.
What is in the context and what fell out of it — a stacked budget with the reply reserve drawn as a segment, and the dropped chunks named with the reason, which is the part nobody ships.
Highlight a passage and hand it to the model. Positioned from the range rect relative to its container, raised by selectionchange so Shift-Arrow works, and every button prevents the mousedown that would collapse the selection.
A suggestion scored with a native meter rather than a styled bar, banded in words as well as numbers, and showing what the model rejected — the fastest way a human catches a bad recommendation.
Read and write scopes separated rather than bundled, writes off by default, a required expiry, and switches that are real checkboxes with role="switch" — so the grant is toggleable by keyboard and announced as on or off.
Forty decisions and an honest way to clear them: bulk approve that cannot reach the high-impact rows, a select-all that says how many it takes, and the auto-approve deadline on every line.
Which agent actions run unattended and which need a signature, stated as a policy the operator can read before anything happens.
What is waiting on a person, ordered by how long it has been waiting rather than when it arrived.
Where a person sits in an automated run, written as four guarantees rather than as a promise that a human is involved.
The published footnotes under an answer — ordered list, real cite and time elements, the claim each source supports, and a staleness warning written in words rather than shown as an amber dot.
Choosing what the assistant may read, with indexing state told honestly — a half-indexed source says so rather than showing a green tick — and the footer stating the effect of the selection, not its cardinality.
The refusal screen: says plainly that nothing was found, shows where it looked and the near-misses below threshold, and offers the two real ways out instead of inventing an answer.
A green tick means the connector works, not that the answers are current. Freshness as an age in words per source, the headline taken from the stalest one, and a failed sync that still says what age it is answering from.
Retrieved sources ranked by how stale they are, because a confident answer from a two-year-old document is the failure people do not catch.
Which sources this answer may draw on, shown as a scope the user can read before asking rather than a citation list after.
How an answer is tied to its sources, and what the system does when it cannot find any.
Copilot-style completion in a plain textarea: a transparent-caret field over a metric-matched mirror, Tab to accept, Escape that actually sticks, and the suggestion announced rather than left as invisible grey text.
The model edits real properties and every change is attributed and reversible — per-property provenance in words, per-property revert, and sliders whose aria-valuetext carries the unit.
Findings the agent surfaced on its own, paged one at a time, each stating its magnitude rather than its direction, dismissible, and carrying a sparkline that is also a sentence for anyone who cannot see it.
The "/" menu inside a document: focus never leaves the editor, every row names what it acts on, and Escape cleans up the trigger text it inserted instead of leaving you to backspace it.
The actions offered on a selection, each saying whether it rewrites in place or opens a draft beside it.
Message threads, streaming states, composers and citations.
Cmd-K dialogs, search results, filter rails and facets.
Top bars, mega menus, mobile drawers and app sidebars.
Dialogs, side panels, bottom sheets and confirmations.
Month grids, availability pickers, booking flows and time slots.
List-edit-delete screens, inline editing, bulk actions and audit rows.