lupy.io
Sections
EN
Download

Letting AI agents work without losing control

Autonomy and oversight do not have to pull against each other. Here is how approval points, plain status and local files keep you in charge of an AI team.

Published: 6 min read

The reason many people hesitate to give an AI real work is not that it might do the work badly. It is that it might do something you did not agree to. Send a message to a client. Spend money. Remove a file. Doing the job well is a quality problem. Doing the wrong thing is a trust problem, and trust problems are the ones people remember.

This post is about how an AI team can run a long job on its own and still leave you in charge. It is a description of the design choices in lupy.io, not a promise that nothing can go wrong.

Two kinds of actions

It helps to split what an agent does into two kinds. The first kind stays inside the work: drawing a screen, drafting a text, collecting prices, summarizing a document. If the result is bad, you throw it away and ask again. Nothing outside the project changed.

The second kind reaches outside: spending money, publishing, messaging someone who is not you, deleting. These are hard or impossible to undo. They deserve a different rule.

A different rule for the second kind

In lupy.io, autopilot runs the job from start to finish, and the second kind of action is designed to stop and ask. The ask arrives in your inbox as a short card with a clear label such as "I need your OK to send the app to the store," and two choices: approve, or not now. Until you answer, the action does not happen.

That is the whole mechanism, and its simplicity is the point. You do not have to watch the screen all day. You only have to answer when a card arrives.

Seeing what is happening

Control also needs visibility. If you cannot tell what the team is doing, an approval card is just a surprise. lupy.io shows work in a few plain ways: task cards that sit in to do, running, waiting or done; a team map that shows who has which job; and live bubbles in which agents write, in their own words, what they are doing right now. One look tells you who has what.

The inbox gathers three things in one place: approvals, questions from agents, and finished results. So there is one place to check rather than six.

Why speed is not the point

It is tempting to judge an AI team by how fast it finishes. For work that touches the outside world, a better measure is how often it stops at the right moment. A fast system that never asks is the one to be nervous about. A system that runs quickly through the safe steps and then waits for you at the risky ones gives you most of the speed and almost none of the worry.

So when you evaluate any tool of this kind, ask two questions. What does it do without asking? And what does it always ask about? The answers tell you more than a speed demo does.

A small example

Imagine you want to send the café app to the store. The agents draw the screens, test them and prepare the version; but right before it goes to the store, the work stops and a card lands in your inbox. The card says what is about to happen and has two buttons: approve, or not now. If you say "not now," nothing is sent.

Meanwhile you are busy with something else. You come when you are ready, read the card, look at the app and decide. There is no time pressure, because waiting is the default behavior.

The small details of visibility

The live bubbles look like a small thing, but they matter for trust. When an agent says "I collected prices from three rivals" or "I am writing the first week of posts," you no longer need to ask what is going on. The information comes to you; you do not go to it.

In the same way, the status on the cards shows who has a job and how long it has been standing. You notice that a job is stuck without asking anyone.

Your files stay with you

Another form of control is where the work physically lives. lupy.io keeps your project files on your Mac.

Your accounts, your terms

The agents run on AI accounts you own, your Claude accounts and your Codex account. That means the usage is under the terms you already agreed to with those providers. How the accounts fit together is covered in its own post.

What approval does not do

It is worth being honest about the limits. An approval card shows you what the team wants to do. It does not guarantee the work behind it is perfect. If the app is going to the store, you should still look at the app. The card makes sure you get to look; it does not look for you.

And approval is only as good as the attention you give it. If you tap approve on every card without reading, you have rebuilt the problem. The design makes the question easy to see; answering it well is still your job.

Setting limits in the request

You can also place limits earlier, in the request itself. Saying "do not send any replies" or "show me the first week before anything is scheduled" shapes the plan so those steps are held back. How to write a request AI agents can finish has examples.

Questions worth asking any tool

Before you hand real work to any AI system, three questions are worth an honest answer. Where do my files live? Who pays for the usage, and how? What does it do without asking me? If the answers are vague, treat that as information.

A reasonable way to start

If you are cautious, start with work that stays inside the project: research, a draft, a report. Read the results, see how the cards and bubbles feel, and move on to bigger jobs when you are comfortable. Trust usually grows from watching a tool behave the same way a few times, not from a promise.

If you want to see the plans and what they include, the pricing page has them, and the FAQ answers the questions people ask most before they start.

In short

Control does not come from the team doing nothing; it comes from the team stopping to ask on actions that touch the outside. Your files staying on your Mac, the accounts being yours and being able to see what is happening are the other legs of this design. None of them promises perfection; all of them keep the decision with you. When you start, begin with jobs that stay inside the project, see how the cards behave, and build your trust from observation, not from a promise. If a job matters to you, do not pass an approval card without reading it; the card is there for exactly that. As a small habit, ask on every card: is this really something I want to allow? The question takes a few seconds and the answer is usually yes; but on the rare occasion it is no, it makes all the difference.

If you want to see the plans and what they include, the pricing page has them, and the FAQ answers the questions people ask most before they start.