The agent demos I watch sell the same fantasy. You type one sentence, close the laptop, and come back to a finished task. Nobody in them gets asked a question, and the agent handles it.

I wanted that too, for about a month. Then I built an agent that actually did it, handed it a real errand, and watched it execute with total confidence. It was wrong, and it never paused to check. That was the day I decided I had the goal backwards.

What I want is an agent that interrupts me at the right moments and stays quiet the rest of the time. For me, those interruptions are most of what makes an agent worth trusting.

The frictionless demo is selling you a liability

Count the questions the agent asks in the demos I’ve seen. It’s usually zero, because the pitch is that it doesn’t need you: you delegate, it executes, you’re free.

That works when the task is reversible and falls apart the second it isn’t. There is a category difference between "draft a reply to this email" and "send this reply to this email." You can throw the draft away. The sent reply lands in someone else’s inbox and stays there.

A frictionless agent treats both the same, because from the inside they are just two tool calls, made with the same confidence and very different consequences.

So the frictionless agent is a fast, tireless worker with no sense of which mistakes it can take back. That makes it a liability you happen to own.

An agent should stop and ask before a step it can’t take back, and stay quiet on the rest.

Whether a step can be undone is the line that matters

Once you stop chasing zero interruptions, the design problem gets concrete. Every action an agent can take falls on one side of a line: can this be undone cheaply, or not?

Drafting is reversible, and so is renaming a file or adding a calendar hold you can delete. Sending an email isn’t, and neither is paying an invoice, deleting the only copy of something, or replying to forty people at once.

The valuable agent acts freely on the reversible side and stops on the other. Reversibility is the signal for when a person should look: you don’t need to supervise something you can undo in a second, and you badly want to supervise the one step that commits.

This is also why "just review everything it did afterward" doesn’t scale. If the agent already sent the email, your review is an autopsy. You are reading the record of a decision that is already out in the world. Move the checkpoint to the moment before the irreversible step and you review one thing, at the one time your judgment can still change the outcome.

The interrupt carries more information than the result

When a well-built agent stops and asks, it also hands you a compressed view of its plan at the point that plan becomes real.

Say you ask it to clean up a folder of files. A frictionless agent deletes what it thinks is junk and reports back that it’s done. A good one comes back with: "I’m about to delete these nine files, keep these four, and here’s why."

That pause is worth more than the cleanup. In one glance you catch the single file it misread, before it’s gone. You didn’t have to audit all thirteen, only one decision, framed by the agent, at the moment it mattered.

That’s a different relationship than autocomplete, which guesses and leaves you to fix it later. A good agent proposes, shows its reasoning at the commit point, and lets you veto. The interruption is how you see what it’s about to do.

When an agent stops to ask, it shows you the one decision that matters while you can still change it.

The quiet matters just as much. An agent that asks about everything is as useless as one that asks about nothing, because you learn to click "yes" without reading and the confirmation becomes theater.

The skill is calibration: silent on the steps you can take back, loud on the few that commit. Get that right and the interruptions stay meaningful, because they are rare and always about something you can’t take back.

Archie stops before anything goes out in your name

Where the "ask me first" boundary sits is a judgment call. Several agents let you move it in their settings, so you could let an agent send routine replies but never touch money, or the reverse. We decided against that for anything sent in your name: in Archie, nothing goes out until you say yes, and no setting turns that asking off.

This is the part I care about at Otian. Archie runs on your own computer, on the AI account you connect, with your files and its memory kept on your computer. What it reads goes to your AI company when it works. So when it pauses before a step you can’t take back, it’s pausing over your actual stuff.

Add-ons change what your agent works on, and the stops stay where they are, whichever add-on suggested the action. A routine can gather and draft on a schedule all day, and its drafts wait for you like any other. A routine running on its own can’t apply a calendar change at all; it reports the change it would make.

That’s the version of "your own agent" I want: it runs on a computer you own, it gets on with the gathering and the drafting, and it stops at the steps that commit. The goal was an agent that needs you at the right moments and leaves you alone the rest of the time.