Agent tasks

An instruction given once, applied to the whole text and re-applied each time you pause typing. Every word it changes is marked and revertible.

Agent tasks: a persistent instruction

The instruction persists

agentically <instruction> _ covers current and future text.

you type

I rite some text witth typos agentically correct spelling _

you get

I write some text with typos

Unlike a transform, which runs once, the instruction stays active: text you add later is corrected on the next pause, without re-issuing the request.

Every change is marked and revertible

Changed words are dimmed, with your original as an alternative.

Corrected words are flagged the way spelling suggestions are. Stepping to one shows your original as the other cycle stop, and Ctrl+Alt+↓ restores it.

Agent tasks change text without a keypress per change. The constraint that makes that acceptable: every change is visible, and every change is reversible.

Runs on a typing pause

The task runs after a pause in typing, not per keystroke. If a result arrives while you are typing, the parts you changed in the meantime are kept and the task's version of those parts is dropped. Your typing always takes precedence.

Suited to rules, not one-off requests

Works wellBecause
correct spellingit applies the same way to every sentence
translate to frenchevery new sentence gets the same treatment
use british spellingyou would otherwise have to remember it every time

A one-off change is better served by a transform. Agent tasks fit rules you would otherwise re-check manually at the end. Typing stop task _ ends the instruction; current task _ shows what is armed.