When an agent needs you
The chat waits as long as it takes, the one that stopped says so, and an answer you tapped on a dead connection still arrives.
Agents stop. They stop to ask which of two things you meant, to ask permission
for something unusual, or to show you two options and let you pick.
The chat waits
While a question is open, that chat is not tidied up. The process
stays, and nothing about the turn is thrown away — an agent that asked you
something at 3am is still holding the question at 9.
It is sized for a night's sleep rather than for ever. With the tab open it waits
as long as you leave it; with nobody looking, a question is held for twelve
hours, and the machine lets the oldest one go if more than eight chats
are sitting on an unanswered question at once, or if it is running short of room
for new work. What goes is the agent's process — the thread is still there, and
the question is one you ask again.
You can see which one stopped
You do not have to open four chats to find the one that is waiting. A
thread with a question open says so where you were already looking — on the
list of chats, on its chip in the shelf, and on the
task it was started against.
So "is anything waiting on me?" is answered by the first screen, at a glance.
[!NOTE]
A question keeps for twelve hours with nobody looking, so arriving three hours
late costs you nothing.
The answer gets there
Tap an answer on a train, go into a tunnel, come out again: the answer is
delivered when the connection comes back. You do not have to say it twice.
The same is true the other way. Close the tab while a question is open and it is
still there when you come back.
What a question looks like
- A permission card — "may I run this?" — with what it wants to do written
out.
- A question with options — tap one. Options are one tap; a paragraph typed
with your thumbs is not, so a well-behaved agent gives you choices.
- A picture to choose between — a page you can actually press, when the
question is about a screen. See
seeing it before it is built.
Queueing instead of answering
You do not have to wait for a turn to finish to say the next thing. Type it and
it queues; the agent picks it up when it comes up for air. Useful for "…and when
you're done, run the tests".
If you would rather not be asked
Give the chat a permission mode with more rope. See
models, effort and permission. The
trade-off is yours: fewer interruptions, less say in what happens.