A note from us: the agent in this story was one we built for our own use. Archie can’t move or rename files on your computer, and in Archie, nothing goes out in your name without your yes.
I said "I'm thinking about moving all my invoices into one folder" to my agent, the way you'd say it to a person who would nod and wait. It didn't wait. It moved them.
Nothing broke. The files were fine, the new folder was sensible, and putting them back took a couple of minutes. But I hadn't decided anything yet. I was doing the thing you do right before you decide: saying the option out loud to hear whether it sounds right in your own ears. My agent heard a completed request and completed it.
You may have felt a smaller version of this in a chat app. You type "maybe I should restructure this whole document" as a thought, half hoping it will talk you out of it, and it returns the entire restructured document. You never said go. It went.
The register you use with people doesn't survive the trip
When you talk to a person, a lot of what you mean lives in how you say it. "Maybe we should." "I wonder if." "What if I just." Those openings are signals, and a colleague reads them without effort. They tell the listener you are feeling something out, so the colleague answers the idea and leaves the doing alone until you say the word.
The model doesn’t read those signals the way a person does. Tentativeness is a social convention, and the model treats your sentence as a prompt to fulfill.
Its entire job is to produce the most helpful next thing, and to almost any input shaped like "I could do X," the most helpful next thing is X, finished. The hedge in front barely registers. "Maybe rename these" and "rename these" land in nearly the same place.
A model hears “maybe rename these” and “rename these” as nearly the same request.
In a chat window the cost of this is small, because all it can do is write text. Worst case you get a wall of output you didn’t want, and you ignore it.
The moment an agent can touch your files, your folders and your accounts, the same confusion changes weight. Now the difference between hearing a musing and hearing a command is the difference between a suggestion and forty files with new names. It’s the same misread with a much larger blast radius.
The tentative voice is where your thinking lives
I only noticed this after it happened to me a few times. The hedged sentence was a different act from a command altogether.
When I say "maybe I'll move the invoices," I am holding the decision open for one more second so I can look at it. I am checking whether I actually believe it before I commit.
That pause is the work. It’s the part where I catch that half my invoices are still tied to reports that point at the old path. That’s exactly the kind of thing you see in the two seconds between saying an idea and doing it.
An agent that turns the maybe into action instantly risks the wrong move, and it also deletes the pause I was using to think. It hands me the result of a decision I hadn’t finished making. Then instead of deciding whether to do the thing, I’m reviewing something already done, which is a worse position, because now there is cleanup attached to changing my mind.
Draw the line between talking and doing
Making the agent timid would be the wrong fix, and so would training yourself to talk to it like a legal contract, weighing every verb so you never accidentally trip it. Talking loosely is how you think, and an assistant that punishes you for thinking out loud is one you will stop talking to.
The line belongs between talking and doing. Chat should stay cheap: say anything, float anything, reason in circles, with no consequences.
Anything that reaches out and changes something real should have to check with me first, every time. That covers renaming, sending, deleting, overwriting, and moving money or files. "Maybe" carries information the agent cannot reliably decode, and when you can’t read a signal, the right move is to ask instead of guess.
In Archie that line doesn’t move for anything sent in your name: nothing goes out until you say yes, and no setting turns that asking off. Say "maybe tell Dana I can’t make Thursday" and your agent drafts the reply, then leaves it as a draft until you send it. Some other agents let you move the line in their settings.
I still think out loud to my agent. It can hear every half-formed idea I have, and it can’t send a reply or cancel a meeting on one until I actually say the word.