The reservation confirmation popped into my inbox while I was still in my pajamas. It promised a Midtown table for two at 6 p.m. — and a bill that made my eyes narrow.
I tried to give Muse, Meta’s new personal AI agent personified by a cartoonish avatar called Jolly, a fair shot. You’ve seen the headlines: Mark Zuckerberg has put Muse at the center of Meta’s pitch as it pours roughly $145 billion ($145 billion; €133 billion) into AI. OpenAI shipped a lookalike product called dots. Big bets, big mascots, and a lot of hope that these agents can save us time.
My first test: it booked a restaurant that didn’t match my budget. Housekeeping tasks with Jolly
I asked Muse to find a reasonably priced spot in Midtown and reserve a 6 p.m. table for two. It picked a Mediterranean place, sent a reservation, and added the confirmation to my calendar—impressive on the surface.
Then I opened the menu. Prices were far above the range I’d given. Conversation with Muse slowed; scrolling through its suggestions felt like chasing a clue someone had left half-buried in a drawer. At that moment Jolly felt less like an assistant and more like a well-meaning but clumsy bellhop.
That small victory—seeing a confirmation arrive without opening a browser—explains the appeal. But the mismatch between instruction and outcome is where trust frays. If you hand an agent control over your apps, you expect it to respect the guardrails you set.
I gave it receipts and flight details; it wrote polite emails but didn’t finish the job. Refund requests and limits
After a pair of late flights, Muse drafted refund messages from screenshots I provided. The letters were good: concise, firm, courteous.
But buried in the draft was a note saying the airline won’t handle refunds via email and that I needed to paste the text into their portal. Muse had done the creative work but stopped short of the manual heavy lifting, leaving the final push to me. That stop-start rhythm is exhausting when you expect seamless help.
Can Muse book flights for me?
Not entirely. When I asked Muse to shop for holiday flights and to connect to Expedia, it offered itineraries but couldn’t complete checkout. It told me it couldn’t access some sites directly, and then stalled on pages protected by CAPTCHAs or anti-bot rules.
United’s terms explicitly bar bots from scraping or logging into their site, and Amazon recently barred Muse from interacting with its store. When those gates rise, Muse will prompt you to take manual browser control—and sometimes you can’t even log in from the embedded window. For me, a task that might have taken 20 minutes on Expedia turned into a fragmented back-and-forth that drained more time than it saved.
I tried to use miles and hit roadblocks on airline sites. Holiday travel booking
I checked United for award space and found prices climbing north of $1,000 ($1,000; €920) for routes I wanted. Muse couldn’t help pull my miles or complete a booking because the airline blocks automated access.
That reality matters. Agents promise to be an extra pair of hands, but if major platforms treat them like interlopers, their reach is clipped. The outcome is an agent that is part guide, part mirror—you see options, but you still do the last mile.
Is Muse safe with my data?
Meta is blunt in its signup flow: Muse can make mistakes and it asks for access to your email and calendar to be useful. I tried the free tier—Muse includes a weekly token limit, and Meta sells higher-usage plans at $20 and $100 per month ($20; €18) and ($100; €92) respectively.
That trade-off—privacy for convenience—is the core decision you must make. If you grant wide access across Gmail, Google Calendar and other apps, the agent learns faster and will likely become more helpful. But granting that access is also a bet on Meta’s security practices and the wider ecosystem, which has already seen incidents like the Hugging Face breach and firms explicitly blocking agent access.
It saved me a few minutes, but not my time. Not yet the age of agents
Across these experiments, Muse often felt like a Swiss Army knife with blunt tools: promising versatility, but lacking the precision to finish the task without human intervention.
I don’t think this tech is useless. Give an agent persistent access to the right apps, and it’ll learn your patterns and remove friction. But today those gains are intermittent. Slow responses, partial automation, and hard stops at CAPTCHAs or anti-bot policies turn what should be frictionless help into a new kind of friction.
Companies are making massive investments—Meta spending billions and positioning Muse as a centerpiece—and you can see why investors care. Yet those bets are just that: speculative plays on user behavior and partner cooperation. For now, you’re likely to keep your habits: booking flights on Expedia, logging into United, and pasting refund text into airline portals yourself.
You can imagine a near future when agents do routine tasks without supervision. But that future depends on platforms, regulations, and security practices aligning with the agent vision. If those players keep treating agents like intruders, how fast do you think we’ll hand over our inboxes and calendars to cartoon assistants?