This week I opened the news and ran into the launch of Meta Muse, Meta’s new personal AI agent, and I confess I spent a good while staring at the screen thinking: “would I let it?”. Because Muse is not just another chatbot that answers questions. It reads your email, handles your calendar, fills out forms, makes purchases and, believe it or not, even negotiates bill discounts for you. On its own.
And then a phrase popped into my head, one we know very well here in Brazil: power of attorney. Hang on, I’ll explain.
What exactly is Meta Muse?
First, the facts. Muse was launched on September 8, for now only in the United States and only for people over 18. You can use it on the web, through the app (iOS and Android) and, look at that, right inside WhatsApp.
The idea is an AI personal assistant that works for you: you tell it a goal (“organize the end-of-year trip”, for example), it builds a plan and starts executing on its own. According to TechCrunch’s coverage, it sends emails, turns a recipe into a shopping list, sends invitations, completes sign-ups and finishes purchases using Stripe Link (with Shop Pay and 1Password coming soon).
There is a free plan with usage limits, plus subscriptions at US$ 20 and US$ 100 per month for those who want to run more tasks. Zuckerberg described the vision as agents working “24/7 on your behalf”. Ambitious? You bet.
Hang on, I’ll explain: this is a power of attorney
Think about it with me. When you need someone to handle something at the notary or the bank on your behalf, you sign a power of attorney. And nobody in their right mind signs a full power of attorney for the first acquaintance who shows up, right? We grant specific powers: only for that matter, only for that period.
An AI agent is exactly that: a digital proxy. The difference between it and a regular chatbot is the same difference between someone who gives you advice and someone who signs papers for you. I have already written here about the difference between agents and automations (in Portuguese), and Muse is the most concrete example so far of this idea reaching everyone’s life, not just tech people’s.
Power of attorney goes to those who have earned it. Keep that sentence, because it sums up everything that comes next.
Meta’s AI agent with your house keys: where the risk lives
Here comes the part that makes me uneasy. Autonomous agents still make mistakes, and they make them while acting. Axios reported the case of a researcher at Meta itself who had emails deleted by an agent from another company, and of services that have already canceled user accounts because they detected bots acting on their behalf. In other words: the damage from an agent’s mistake is not a silly answer on the screen, it is an action in the real world.
And there is the other side: we are talking about Meta, a company with a privacy track record that includes a US$ 5 billion fine and the Cambridge Analytica case. I am not saying things stay the same forever (companies change, it happens rs), but the distrust here is not paranoia, it is memory.
To be fair, Meta thought about this. Muse runs on a dedicated virtual machine, with a system called Sentinel that decides what the agent can do alone, what gets blocked and what requires your approval. Permissions separate reading from writing and can have an expiration date. On paper, it is a well-drafted power of attorney. Reading is one thing, signing is another! The question is whether, in everyday life, we will have the patience to set this up properly or will just go around clicking “allow everything”.
What I would check before connecting anything
If Muse (or any similar agent) knocked on my door today, this would be my playbook. First: start with read-only access. Let the agent see your calendar and suggest things, without being able to touch anything. Trust me, leave writing for later.
Second: no connecting payments in the first week. Cards and bank accounts are the last frontier, and only after the agent has proven itself on harmless tasks. Third: expiration dates on permissions, whenever possible. An eternal power of attorney is an invitation to headaches.
Fourth: turn off the use of your conversations to train the company’s models. In Muse, this sharing comes enabled by default and can be switched off in the settings (Meta promises a confidential version by the end of the year, in which not even Meta sees your activity). And fifth: review what the agent did every week. It is the same habit as checking your credit card statement. Boring? Yes. Necessary? Also yes.
Start small, always. Autonomy is something you grant little by little, just like with people.
And what about Brazil?
For now Muse has not arrived here, but notice the detail: it works inside WhatsApp, the app Brazilians open before their morning coffee. When this kind of agent lands in Brazil, adoption will not be slow and gradual, it will happen overnight (hello, novelty lovers rs).
That is why I think it is worth preparing now. Understanding what an agent is, what read and write permissions are, what connecting an account means. Whoever learns this today, calmly, will not sign a full power of attorney in a panic tomorrow.
Frequently asked questions
Is Meta Muse already available in Brazil?
No. For now it is only in the United States, for people over 18, via web, app and WhatsApp. Meta has not announced a date for other countries.
Is Meta Muse free?
There is a free plan with task limits, which according to Meta covers most people. Those who need more can subscribe to the US$ 20 or US$ 100 monthly plans.
Does Muse have access to my card and my passwords?
Meta says it does not store your passwords or payment data: purchases go through services like Stripe Link, and you choose what to connect, with separate read and write permissions. Even so, my suggestion is to enable payments only after a lot of trust has been built.
What is the difference between Muse and a chatbot like ChatGPT?
The chatbot answers you; the agent acts for you. Muse opens a browser, fills out forms, sends emails and makes purchases. That autonomy is what makes it useful and, at the same time, riskier.
An AI agent is a proxy: specific powers, short terms and trust built little by little.
Cheers, Fellipe Soares