Personal AI agents, assistants that work in the background, browse on your behalf and message you first, have gone from demo-ware to something people actually use. WIRED’s Reece Rogers set up six of them on one iPhone and ran five tests over 24 hours: a morning news briefing, a group lunch order, spotting a calendar conflict, planning a trip itinerary and a late-night “help me” check-in. Each attempt was scored from 1 to 5.
The lineup and prices (as reported)
- OpenAI Dots: inside ChatGPT, $100 a month, the priciest tested.
- Meta Muse: free, and topping app download charts since its September release.
- Anthropic Claude Duet: $20 a month, a waitlisted research preview that launched the night before testing.
- Google Gemini Spark: a beta in the Gemini app, $20 a month.
- xAI Grok Bot: $30 a month, in its own app.
- Instinct: free with an invite code, and it works over text messages instead of an app.
How they did
- Best overall: Dots. It performed best across the tests, though WIRED suggests most people wait for a cheaper plan.
- Most accessible: Muse. A close second, free and the easiest to set up, though the reviewer flags Meta’s track record on data.
- Best interface: Claude Duet. Middle of the pack on scores, but its app was the easiest for following ongoing tasks.
- Most annoying: Instinct. The text-only design left the tester feeling disconnected from what it was doing.
Two details stand out for work use. Only two of the six caught a conflict between an email reminder and a calendar meeting before it happened. And on the travel test, all six suggested the same route, a reminder that agents can push everyone toward the same answers.
A checklist before your team adopts an agent
- Start with limited access. WIRED connected only a personal calendar and Gmail. Do the same in a pilot before linking work accounts or payment methods.
- Watch where it posts. WIRED cites a user whose agent shared his bank balance in a work Slack. Restrict which channels an agent can write to.
- Check how usage is billed. Proactive agents that run tasks unprompted may count against your plan’s limits.
- Test the tasks you actually need. Scores varied a lot by task, so run your own short trial.
- Ask about sharing and storage. See our piece on Muse’s share-link exposure for why this matters.
Based on reporting by WIRED. Prices and features are as WIRED reported and may change.
