Wired Tested 6 AI Agents in a Day: Dots Leads, Gemini Spark Fails at Lunch

A year ago, Wired reporter Reece saw AI agents as a useless demo for slide decks. Today, always-on agents are moving beyond Silicon Valley because they have become more reliable: they finish tasks more consistently and handle a browser with more confidence. The day before, Reece set up six agents on his iPhone, ran them through five tasks over a day and scored each attempt from 1 to 5. Three of the tests already show how the field stacks up.

Who took part and what it costs

The agents work like a group chat: a single message feed, plus a virtual browser that acts as their own computer for getting online. Unlike a chatbot, they message you first when they see a chance to help. The test included:

• Meta Muse — free, launched in September and has topped smartphone download charts ever since.
• OpenAI Dots — built into ChatGPT, $100 a month. The most expensive of the lot.
• Instinct — free, but the user needs an invite code. It runs over ordinary SMS, with no app.
• Anthropic Claude Duet — $20 a month and the newest: a research preview available via waitlist, launched the day before the tests.
• Google Gemini Spark — a beta inside the Gemini app, $20 a month.
• xAI Grok Bot — a standalone app, $30 a month.

OpenClaw did not make the list. It kicked off the wave of agents earlier this year, but it is too hard for an ordinary person to set up. The author promises to add Hermes from Nous Research and Poke to future rankings.

Briefing and lunch: where the agents stumble

The first task was to prepare a pitch and a headline for Wired's editorial meeting. Reece asked the agents to spend 30 minutes on deep research. Only Muse and Dots worked for that long; the rest delivered almost instantly or after a couple of minutes. Gemini Spark offered the best idea (5/5): Walmart is opening a platform for third-party agents, while Amazon has shut Muse out. Claude Duet got 4 for a topic about PIN-protected ChatGPT chats, and Muse got 3. Dots got 2 for an "evergreen" piece about SHIFT, an anti-doomscrolling app. Grok Bot and Instinct got 1 point each: the former pitched old news about Muse's "basic taste," the latter a vague experiment with social media feeds.

The second task was to order lunch for two. The boss is vegan and wanted Vietnamese food, Reece needed meat and caffeine, and the restaurant had to be near the office.

• Dots — 5/5. Picked DragonEats on its own and ordered through the restaurant's website: bánh mì with roast pork, Vietnamese iced coffee, teriyaki tofu rolls. Total: $30.57 before tax.
• Muse — 4/5. DragonEats via DoorDash, $42.66 for the items.
• Grok Bot — 4/5. Freshroll via DoorDash, $43.40.
• Claude Duet — 3/5. Freshroll; it asked whether to use DoorDash or Uber Eats. Reece chose Uber Eats, $49.25.
• Instinct — 2/5. The cart resembled Claude Duet's but cost $4 more: $53.73.
• Gemini Spark — 1/5. The boss got a single roll for $5.85, with a subtotal of $22.55. The only real failure in this test.

Proactivity: the 14:00 conflict

Reece created a calendar meeting with a source for 14:00 and emailed himself a note that he needed to call his mother, who was being discharged from the hospital, at that time. The task was to connect the email to the calendar and warn him in advance.

• Dots — 5/5. Noticed right away and suggested calling before the meeting or after 15:00, taking the discharge time into account.
• Claude Duet — 4/5. Warned him several hours ahead but offered no alternatives.
• Grok Bot, Instinct, Muse — 2/5 each. Missed the conflict, but admitted the mistake and offered to check email against the calendar in the future.
• Gemini Spark — 1/5. Did not find the overlap and did not suggest how to avoid a repeat.

Only two succeeded, fewer than the author expected.

How to choose an agent right now

After three tasks, Dots (12 points out of 15) and Claude Duet (11) lead, followed by Muse (9), Grok Bot and Gemini Spark (7 each) and Instinct (5). But this arithmetic does not capture cost and risk. Dots costs five times as much as Claude Duet and Gemini Spark, while Muse is free.

Agents need access to email, calendar and other personal data, and permission mistakes are expensive: a user on X connected agents to a bank account and a work Slack, after which one of them posted his balance in front of colleagues. The practical advice: start with calendar and email, and do not give agents access to banks or corporate chats. Check carts and routes before paying, especially if the agent worked too fast.

In the test, Reece connected only his personal Google Calendar and Gmail, which is a sensible way to start.