Insights · Agents

Before You Let an AI Agent onto Your Computer

Personal AI agents are moving onto the computers people already use. Seven questions worth asking before one acts in your name.

Until recently, most of the personal AI agents in the news lived somewhere else. You messaged them, they did their work on a computer in someone else's data centre, and they reached your email and calendar through whatever access you handed over.

That's changing. More of them are heading for the computer you already use: the one with your files, your signed-in apps and your open browser tabs. An agent that works where you work can do far more for you. It also changes what you should ask before you let it in. The question is no longer where it runs. It's what it does there when you're not watching.

These are the seven questions we'd ask of any agent, our own included.

1. What can it reach?

Start with the list. Which apps, folders, accounts and devices can it touch, and can you see that list in one place? An agent can only misuse what it can reach, so a good answer is specific: these apps, these folders, this account. "Whatever you need it to do" is an answer too, just not a reassuring one. Give it the least it needs for the job you have in mind, and add more once you've seen how it behaves.

2. Does it ask before it acts in your name, or does it decide when to ask?

This is the one that matters most, and the difference is easy to miss. Some agents ask when they judge a step to be risky. Others ask by default before anything goes out in your name, and stop asking only where you've told them they can. The first leaves the judgement with the agent. The second keeps it with you.

3. What happens if you don't answer?

An agent that asks is only as careful as what it does when you don't reply. If it waits a while and then goes ahead anyway, the question was a formality. The safe default is the boring one: no answer means no.

4. Where do your logins live?

To act for you, an agent needs a way into your accounts, so ask how it gets one. Some hold your access on their own servers and act from there. Others work through the apps you're already signed in to, so your logins stay on your own machine. Either can be done carefully, but you should know which one you're agreeing to.

5. Who else can make it act?

If more than one agent works for you, ask what each can reach through the others. Access you granted for one job can quietly become available to another. And if your computer sits in a shared room, a request from someone else is still a request. A careful agent knows whose request it is acting on, and keeps each person's access for that person.

6. Can you check what it did, or only ask it?

When something goes wrong, the natural move is to ask the agent what happened. Its answer will be fluent and confident whether or not it's accurate, because that's how these systems talk. What you want is a record: what it was asked to do, what it actually did, and when, written down as it happened and readable by you. An explanation is not a record. We've written more about this in Nobody Should Have to Take an AI's Word for It.

7. What does it learn from you, and where does that go?

An agent on your computer will see a great deal: documents, messages, the shape of your week. Ask whether any of it is used to train AI models, whether that's switched on by default, and whether deleting a conversation also removes what the agent learned from it. The answers vary more than you'd expect, and they tend to live in the help pages rather than the adverts.

How we answer them

We built Morgan to work on your own Windows PC, in the apps you're already signed in to, so your logins stay on your machine. By default it asks before it sends or submits anything in your name, and no answer means no. It knows your voice, tells you apart from other people in the room, and only answers when you're talking to it. Every action it takes is written to a record you can read afterwards. And your conversations are never used to train any AI model, ours or anyone else's.

Hold us to those answers the same way you'd hold anyone else. Our private beta is full, and we're building toward launch. If that's the kind of agent you'd want on your computer, you're welcome to join the waitlist.

Related reading: what makes an AI trustworthy, and what "your data isn't the product" really means.