Always-on AI agents are becoming easier to use, but choosing one means weighing more than how impressive its answers sound. A WIRED reporter spent a day testing six agents on practical tasks, including preparing a tech news briefing and pitching a headline. The results pointed to real improvements in task follow-through, alongside limits that matter when software can access personal information.
What makes an AI agent different?
An AI agent can be treated like a digital assistant you message to handle tasks. The examples in the test included reviewing email for important messages, tracking flight deals and helping with shopping decisions.
Unlike a typical chatbot that waits for each prompt, these agents can be more proactive. They may send a message when they detect a possible opportunity to help. Conversations often happen in one thread, and agents can use a virtual browser to reach information on the internet.
That combination changes the user experience: the software is expected to carry work forward, not just respond to a question. It also makes instruction-following and browser control central to whether an agent is useful.
Six services, with different costs and access
The reporter set up six agents on an iPhone: Meta’s Muse, OpenAI’s Dots, Instinct, Anthropic’s Claude Duet, Google’s Gemini Spark and xAI’s Grok Bot. They varied in price, availability and where people use them.
Muse and Instinct were free; Instinct required an invite code and worked through text messages. Dots cost $100 a month and were built into ChatGPT. Gemini Spark was a beta in the Gemini app for $20 a month, while Grok Bot had its own app and cost $30 a month. Claude Duet cost $20 a month and had just launched as a wait-listed research preview.
The test did not include OpenClaw, which the article describes as too clunky for an average person to set up. The reporter also pointed to Hermes by Nous Research and Poke as smaller services that might be considered in a future comparison.
Following directions mattered as much as the pitch
To compare the agents, the reporter created five tests and scored each attempt from 1 to 5. The exercises took place over 24 hours and were designed to examine proactivity, instruction-following and browsing ability.
For the morning briefing task, each agent received a detailed prompt about the reporter’s coverage areas and how the resulting pitch would be used. The instructions asked the agents to spend 30 minutes on deep research. Only Muse and Dots continued researching for most of that period. The others returned a pitch immediately or after a few minutes.
That difference was a useful signal: an agent’s ability to respect the process requested by a user is part of the task, even before anyone judges the final answer. Yet persistence did not guarantee an appealing result. The six pitches did not impress the reporter’s colleagues in an editorial meeting.
Gemini Spark produced the strongest angle, connecting Walmart’s efforts to let third-party agents use its shopping platform with Amazon’s decision to block Muse. The outcome suggests agents can help surface a possible story direction, but their output may still need a human audience to judge whether it is compelling.
Personal access brings trust questions
The test was limited. The reporter shared a personal Google Calendar and Gmail account with the agents, but did not expose them to all personal or work data. The article cautions that the software is still new and carries privacy and security concerns.
Agents may become more helpful when they can draw on email, calendars and other personal context. That access also asks users to trust the service with information that could be sensitive. The article describes an X user who said an agent connected to a bank account and work Slack posted their account balance where colleagues could see it.
That example underlines why convenience alone is not enough to evaluate an AI agent. People should consider what information they connect and what the software might do with it. The reporter’s one-day exercise offers a snapshot rather than a comprehensive verdict: the tools can improve as they receive more information about a person and their preferences, and the test covered only a narrow range of uses.
For now, the clearest takeaway is that AI agents have become more capable at following through and operating a browser, but mistakes remain possible. A useful comparison has to consider the task, the price, how consistently the agent follows instructions and the level of access a user is comfortable granting.