The promise of AI agents

Meta's Muse app promises to automate Facebook Marketplace tasks, sort emails, manage calendars and buy groceries.

A crowded market emerges

Muse leads with more than eight million downloads in its first month. OpenAI announced Dots, X has Grok Bot, and Instinct operates in iMessage or WhatsApp. All promise to remove friction from busy lives.

Testing the agents

Adam Clark Estes, a senior technology correspondent at Vox, spent two weeks using AI agents.

The privacy price

To work effectively, AI agents need access to Gmail, calendars, credit cards and healthcare portals.This means handing tech companies vast amounts of personal information.

The blank slate problem

People don't know how to start with AI agents because they don't understand the tools' capabilities or what would help.

“I feel like I'm not gaining a lot of time because I can't completely trust the system.”

Start small, build trust

Julie Zhuo, founder of the AI company Sundial and an early Facebook employee, advised starting with smaller tasks. Once you see success, you develop more trust that the agent can handle bigger things.

“In the beginning, it's very hard to trust that it's going to just magically do a really big thing.”

The notification trap

Muse sends constant push notifications prompting actions. It notified Estes about apple-picking, a podcast show, a Halloween party he wasn't invited to, and evening digests. Turning off notifications means losing friction reduction opportunities.

When agents work best

Agents worked better when Estes drove the conversation.Adding calendar events, making to-do lists, being reminded to reply to emails: these simple administrative chores saved real time.Other journalists reported similar success with basic productivity tasks.

The productivity paradox

Allison Johnson at The Verge noted that productivity hacks often don't remove tasks from your plate. They invite more tasks. Agents may save time on individual chores but create new work overall.

“the dirty secret of productivity hacks is that doing more stuff often doesn't feel like things are being removed from your plate; somehow it just invites more tasks.”

Unpredictable consequences

Despite having access to vast personal information, AI agents can misunderstand instructions and cause real damage.Matt Robb gave Muse access to his Facebook Marketplace account.

The ironies of automation

Psychologist Lisanne Bainbridge identified the ironies of automation in nineteen eighty-three. More advanced automation leads to more consequential errors. The more humans rely on automation, the worse they get at catching those errors.

Cognitive offloading risks

Vera Liao, an electrical engineering and computer science professor at the University of Michigan, explained that people consistently accept wrong AI outputs even when they could have gotten it right themselves.

“Even in simple tasks like judging whether AI output is wrong, people consistently accept wrong AI outputs even if they could have gotten it right on their own.”

One mistake cascades

Estes audibly gasped when Instinct sent his email replies without asking.Looking back, he could see why the agent interpreted his instructions that way.

Creating more slop

Pattie Maes warned that outsourcing tasks to agents creates more work for human service workers who must interact with flawed AI.The world risks becoming built only for agents to interact with services, not for humans.

“There's just more and more slop.”

The optimists push back

Matt Van Horn, co-founder of the ridesharing service that became Lyft, calls Instinct magical and pushes friends and family to use it.His wife sent a parking ticket photo to Instinct with no description.

“I've been pushy with my mom and my wife, and saying, 'Hey, you should sign up for this. It's magical.'”

When agents fail in the real world

Van Horn asked Muse to reserve a karaoke bar room.The karaoke place refused to talk to an AI and asked for Matt the person.Muse failed.

“The karaoke place said, 'Are you an AI?' 'I don't want to talk to an AI. Please connect me to Matt, the person.'”

Vitamins, not painkillers

Julie Zhuo planned a college reunion party using agents, tracking emails and sending invitations.But she noted that agentic AI applications read more like vitamins than painkillers.Vitamins offer vague benefits.Painkillers solve big, immediate problems.

“these kinds of applications of agentic AI read more like vitamins than painkillers.”

Choose tasks you can tolerate

Vera Liao advised choosing to use agents only for tasks where you can tolerate the consequences of errors. Estes made a doctor's appointment using Instinct. At best, the times agents worked saved mundane time.

“Choose to use agents for tasks that you can tolerate the consequences of errors.”

Estes steps back

Estes stopped using Muse to avoid creating more emails or alerts.He still texts with Instinct, which messages him flagging emails or offering help.When there's a clear payoff, he replies.Otherwise, he leaves it on read.