I Tested Every Major AI Agent for a Day. Here’s the Best One

A year ago, personal AI agents were mostly demo-ware. WIRED's software reporter put six of them through a 24-hour gauntlet on her iPhone: Meta's Muse, OpenAI's Dots, Instinct, Anthropic's Claude Duet, Google's Gemini Spark and xAI's Grok Bot. The tests were ordinary on purpose: a morning news pitch, a lunch order for two, a calendar clash hidden in an email, a China itinerary and a middle-of-the-night check-in.

The headline result is that these tools now mostly follow through. OpenAI's Dots took best overall, Meta's free Muse came a close second and was easiest to set up, and Claude Duet won on interface despite launching as a research preview the night before. The catch is that the best performer sits behind a $100-a-month paywall, and every agent works best when it can read your email and calendar.

We like this kind of test. It is not a benchmark chart; it is a person trying to get through a Thursday. And by that standard, the category has turned a corner. Agents that keep researching for the time you asked, build a sensible lunch cart for a vegan boss, and flag a conflict between a meeting and a call to a parent leaving the hospital are doing real work.

But look at where they stumbled. Only two of the six caught the calendar conflict, the single most "assistant" task on the list. Most of the others apologized afterward and promised to do better. Gemini Spark put a single spring roll in the cart for a grown adult's lunch. That is the pattern I'm watching: agents are now good enough that people will trust them, and still uneven enough that the trust will sometimes be misplaced.

The privacy cost is the part we don't want lost in the scorecards. The reviewer only connected a personal Gmail and calendar, and she cites a user who said an agent posted his bank balance into a work Slack. These products get better the more of your life you hand them. That is a business model as much as a feature, and it is arriving fastest from companies with advertising businesses and spotty data records. The reviewer herself flagged concerns about Meta even while praising Muse.

There is a quieter downside too. All six agents planned the identical route through China. If millions of people outsource their trips, restaurants and shopping to a handful of models, we get a monoculture of taste, with a few businesses winning every recommendation and everyone else invisible. That is where this leads at scale unless the agents get more diverse in how they search and rank.

Our read: useful, getting better quickly, and worth trying carefully. Start with low-stakes errands, keep money and work accounts walled off, and judge these tools on whether they catch the thing you would have missed, not on how warm their texts sound.

Pricing at the time of the test: Muse is free; Instinct is free but invite-only and runs through text messages; Gemini Spark is a $20-a-month beta; Claude Duet is $20 a month and wait-listed; Grok Bot is $30 a month; Dots requires a $100-a-month ChatGPT tier. OpenClaw, which helped spark the category earlier this year, was left out as too hard for most people to set up.