An inline agent is a continuously-running editing task that lives inside your text field. Start it with agentically <task> _ (say, agentically correct spelling _) and keep typing: every pause re-checks your draft against the task, and any edits land dimmed in place, each one individually revertable.
The whole interface is four phrases
agentically <task> _starts the task (the phrase itself is wiped; the secondary display shows[task: ...]).add task <more> _stacks another concern onto the same task.stop task _stops it. Existing edits stay; revert any of them individually.current task _prints what's running.
Accept by doing nothing, revert with two keys
There is no review pane and no accept button. An edit that lands dimmed is accepted by ignoring it; to reject one, navigate to it with Ctrl+Alt+←/→ and press Ctrl+Alt+↓ to cycle back to what you wrote. Built-in grammar correction rides along on every pass of a running agent.
It never fights you, and never submits
Rewrites are merged against your live buffer, so anything you typed while the model was thinking is never overwritten. And the agent only ever edits your unsent draft: you are always the one who presses Enter. That boundary is the core of the security model.
The agent fires after a 1s pause by default (tunable via agent-debounce-ms) and caches aggressively, so unchanged text costs nothing. The full design story is in the Inline Agents essay.
Related: Fix grammar as you type · GitHub repo