Last year many technology reporters complained that AI assistants performed poorly, but the sector has since shifted. Companies now ship “always-on” agents that can act without prompting, and early adopters report more reliable behavior. While errors still occur, recent versions demonstrate better task completion and web-browser control, prompting a surge in consumer interest beyond niche developer circles.
The author installed six leading agents on a single iPhone for a 24-hour trial. Meta’s Muse, released in September, is free and dominates app-store rankings. Instinct also costs nothing but requires an invitation and operates via text messages. OpenAI’s Dots are bundled inside ChatGPT at $100 per month, while Google’s Gemini Spark charges $20 monthly in beta. xAI’s Grok Bot costs $30 a month, and Anthropic’s Claude Duet, a wait-listed preview, is $20 per month.
To compare capabilities, the tester designed five core scenarios covering proactivity, instruction adherence, and web navigation, assigning each a rating from one to five. The agents received a single thread of messages, mimicking a group-chat interface, and were granted access to the author’s Gmail and calendar only. Over the course of the day the bots attempted tasks such as scanning email, tracking flight deals, and generating a news briefing for a technology editorial meeting.
When asked to spend thirty minutes researching a detailed briefing, only Muse and Dots continued searching for the full interval, while the remaining agents produced headlines within minutes. Gemini Spark’s output received particular notice for framing a story about competing retail platforms, contrasting Walmart’s third-party agent program with Amazon’s legal response. The other bots either halted early or delivered less nuanced angles, suggesting a gap in sustained instruction following.
Security analysts warn that the convenience of personal agents comes with data-exposure risks. One reported user linked an agent to both a banking account and a workplace messaging service, only to see the bot inadvertently post the account balance publicly. Because these assistants rely on continuous access to email, calendars, and sometimes financial information, users must weigh the productivity gains against potential privacy breaches.
The trial demonstrates that free agents can match or exceed paid competitors in certain persistence metrics, yet overall performance remains uneven across tasks. As developers refine browsing automation and instruction handling, future updates may narrow the gap. For now, early adopters should remain cautious, limiting data sharing until privacy safeguards become standard across the rapidly expanding personal-assistant market.