WIRED tests six always-on AI agents, from Meta's Muse to Claude Duet

A WIRED software reporter set out to find which of the new always-on AI agents is best. Before going to bed on a Wednesday, he set up six of them on his iPhone: Meta's Muse, OpenAI's Dots, Instinct, Anthropic's Claude Duet, Google's Gemini Spark and xAI's Grok Bot. He designed five basic tests, ran them over 24 hours to judge proactivity, direction following and browsing, and scored each attempt from 1 to 5, planning to tally the results, crown a winner and give an award for most annoying.

The agents differ in price and form. Muse is free and has dominated the smartphone app download charts since its September release. Instinct is also free, but needs an invite code from an existing user and works through text messages rather than an app. OpenAI's Dots is built inside ChatGPT and costs $100 a month, the most expensive of the group. Google's Gemini Spark is a beta inside the Gemini app at $20 a month. xAI's Grok Bot costs $30 a month and has its own app. Anthropic's Claude Duet costs $20 a month and is the newest, released as a wait-listed research preview the night before the tests. OpenClaw was left out because the author finds it too clunky for the average person to set up; he says smaller players such as Hermes by Nous Research and Poke might join a future ranking.

The author describes an agent as something like a digital assistant you text: it can scan email for missed messages, track vacation flight deals and help with shopping. Unlike a turn-based chatbot, it is more proactive and nudges you first when it spots a chance to help. Messages usually sit in one thread that feels like a group text, and the agents do much of the work by controlling a virtual browser, in effect their own computer. He says agents still make mistakes, but sees significant improvement in following through on tasks and controlling a browser compared with a year ago. He shared only his personal Google Calendar and Gmail with the tools.

Test one was a morning briefing. He sent each agent a detailed prompt about his coverage areas and how the pitch would be used at the weekly WIRED Gear desk meeting, and asked it to spend 30 minutes on deep research. Only Muse and Dots kept researching for most of that time; the others replied instantly or after a couple of minutes. The author reads that as Muse and Dots being a bit better at following instructions in this instance. The pitches themselves did not go over well with his colleagues. Gemini Spark produced the best angle, contrasting Walmart creating ways for third-party agents to use its shopping platform with Amazon recently giving Muse the boot. Colleagues did not seem worried about being scooped by an agent.

Test two was lunch. He added his boss, who is vegan and wanted Vietnamese, to the order, while he wanted something meaty and a lot of caffeine, from a place near the office that he could walk to. The agents all chose similar restaurants, DragonEats or Freshroll, and the team went with DragonEats as a bit tastier and cheaper. Instinct and Claude Duet built carts that were a bit too pricey. Gemini Spark was the only agent that truly failed: it added a single $5 roll for the boss, which the author calls more snack than lunch. The article prints several of the agents' cart summaries without labelling which agent wrote which, with totals ranging from $22.55 to $53.73 before tax and fees.

Test three was schedule conflicts. He created a 2 pm Google Calendar event marked as an important source meeting that cannot be moved, and emailed himself from his work account to his personal one a made-up note that his mother was leaving hospital and he needed to call her at 2 pm. The question was which agents would read both, spot the clash and message him beforehand with alternatives. Only two succeeded, fewer than he expected. The article quotes a mix of replies, some flagging the overlap and suggesting a call before or after the meeting, others admitting they never cross-checked the email against the calendar.

The visible text ends at the start of the fourth test, a vacation itinerary task. The author says he enjoys researching trips and does not want to automate that fully. The final scores and winner are not included in the text available here. He stresses the piece is not an endorsement, since the software is nascent and often tied to privacy and security concerns. He points to an X user who said he connected AI agents to his bank account and work Slack and was mortified when an agent posted how much money he had for colleagues to see.

Key facts

WIRED's reporter set up six always-on agents on his iPhone: Meta's Muse, OpenAI's Dots, Instinct, Anthropic's Claude Duet, Google's Gemini Spark and xAI's Grok Bot, and ran five tests over 24 hours, scoring each attempt 1 to 5.
Prices: Muse and Instinct are free (Instinct needs an invite code), Claude Duet and Gemini Spark cost $20 a month, Grok Bot $30, and OpenAI's Dots $100 a month, the most expensive.
In the morning briefing, only Muse and Dots kept researching for most of the 30 minutes requested; Gemini Spark wrote the best pitch angle, though colleagues were unimpressed.
In the lunch test Gemini Spark was the only agent that truly failed, adding a single $5 roll for the vegan boss; Instinct and Claude Duet built carts that were a bit too pricey.
Only two of the six agents spotted the conflict between a 2 pm calendar meeting and an email reminder to call his mother, and messaged him beforehand.

Why it matters

Always-on personal agents are being released by Meta, OpenAI, Anthropic, Google and xAI, and the author says they now work more reliably than a year ago, with significant improvements in following through on tasks and controlling a browser. A side-by-side test of everyday jobs (a research briefing, a food order, a calendar clash) shows where the differences actually lie. The early results are uneven: instruction following, price discipline and proactivity varied from agent to agent.

Who it affects

Mainly people considering a first AI agent who are unsure which one fits their needs; the author writes for nontechnical users. It also touches the companies racing to attract early adopters, and anyone weighing the price gap, from free (Muse, Instinct) to $100 a month (Dots).

How to use it

The article is a buyer's guide in progress rather than a how-to. Prices to weigh: Muse free; Instinct free via an invite code and text messages; Gemini Spark $20 a month as a beta in the Gemini app; Claude Duet $20 a month as a wait-listed research preview; Grok Bot $30 a month in its own app; Dots $100 a month inside ChatGPT. The author notes the agents get better as you share more about yourself and how you like tasks done, and that they work best when connected to email and calendars.

How solid is it

This is one reporter's hands-on test, not a benchmark: five tasks, one day, one phone, and attempts scored 1 to 5 by the author himself. He says the tests are a limited sliver of what the tools can do. The text available here ends partway through the fourth test, so the final scores, the winner and the most-annoying award are not covered. Which two agents caught the calendar conflict is not stated in the portion available.

Risks and caveats

The author says plainly that this is not an endorsement. The software is nascent and often comes with privacy and security concerns: agents work best with personal data such as email and calendars, which demands trust. He cites an X user who linked agents to his bank account and work Slack, and was mortified when an agent posted how much money he had for colleagues to see. The author himself shared only his personal Google Calendar and Gmail. Agents also still make mistakes, such as missing a clash that sat in an email rather than on the calendar.