For the last two years, the interesting work in agent design has been happening at enterprises, quietly. How do you let an agent read a Salesforce field without it also being able to read the whole database. How do you let it send a WhatsApp on behalf of a customer without accidentally spamming them. How do you tell the user, in the moment, what the agent is about to do — before it does it.
That work has been invisible because enterprises don't publish design specs. And it's been slow because the audiences are small and the review cycles are long.
Meta Muse changed that in a week.
One billion consumers, all learning what an agent is at the same time
Muse doesn't just add a chat window to Instagram. It gives every Meta account a personal agent that can act — book, browse, remember, cross-app. The rollout is measured in hundreds of millions of users, not thousands. And every one of those users is about to learn, usually the hard way, that letting an agent act on your behalf is a fundamentally different relationship than talking to a chatbot.
The friction shows up immediately. Every screenshot circulating this week has the same shape: the agent did the thing, but not quite the thing. Booked, but the wrong table. Ordered, but the wrong size. Remembered, but the wrong context.
Some of that is model error. Most of it is the consent-UX problem.
The consent-UX problem, briefly
When a human takes an action for you, you get context for free. The travel agent asks clarifying questions. The waiter reads the room. The friend booking your table knows you hate window seats. Agents get none of that — and the interface for filling the gap has been the open research problem of the last two years.
In enterprise deployments the answer has been mostly explicit. Confirm before send. Show the tool call. Preview the write. Ask again if the confidence score drops. It works, but it's ugly, and it kills the magic. In consumer there is no magic to kill if you don't fix it — the user just quietly stops trusting the agent, and the product dies of a thousand small disappointments.
What changes now
Three things.
One: the enterprise patterns get exposed to public scrutiny, fast. Design conventions that were reasonable inside a controlled workflow — approval dialogs, hard-coded scopes, safety rails — get tested by users who never signed a change-management doc. Some patterns survive. Some don't.
Two: the market for consent-UX design becomes real. Until this week it was a niche discipline inside a handful of firms. Starting now, every consumer agent product has to hire for it, and the teams that built the taste on internal tools will find themselves suddenly employable in ways they weren't.
Three: enterprise buyers watching Muse land badly will get nervous. They should. Their agents run in higher-stakes environments — the shipment worth $80,000, the wire transfer, the medical booking — and if the biggest company in the world can't get consent right in front of a billion consumers, the CISO who was already skeptical about your internal agent will find their skepticism validated.
Where we've been shipping this
For GWC's WhatsApp voice agents the consent surface is intentional. The customer says what they want. The agent confirms what it is about to do, in the same voice channel, before doing it. If confidence is low it asks. If the action is irreversible it asks twice. The pattern is simple and it works — but only because we designed the agent around it from the start rather than bolting it on after a demo.
That's the piece Muse skipped. And it's the piece every serious agent product now has to think about, whether they wanted to or not.
What to do this week
- ›If you're shipping a consumer agent: pull three of your own users into a room and watch them use it. Every hesitation is a consent gap.
- ›If you're shipping an enterprise agent: your CISO just got new ammunition. Get ahead of the conversation before they bring it up.
- ›If you're a product leader watching from the sidelines: the design roles you'll need in six months don't exist yet. Start looking now.
The agents are coming for every product. The trust is not automatic. The teams that figure out the consent-UX problem win the next decade.