Key points
- A Wired reporter ran six always-on AI agents through five tasks in 24 hours, scoring each from 1 to 5.
- Only Meta's Muse and OpenAI's Dots kept researching for most of a requested 30-minute deep-research window; the rest answered within minutes.
- Just two agents flagged a planted conflict between a 2 pm meeting and a reminder to call his mother, fewer than the reporter expected.
- All six proposed the same China route for a two-week trip, raising concerns that travelers get steered to the same places.
Why it matters for marketers: Agents now handle real errands, but uneven instruction-following, shallow research and identical recommendations mean marketers should expect agent-driven shopping and travel to funnel people toward the same brands and places.
Wired's software reporter spent a day running six always-on personal agents on his phone, from Meta, OpenAI, Anthropic, Google, xAI and the startup Instinct. He gave each five tasks covering proactivity, following directions and browsing, and scored every attempt from 1 to 5. His overall view: agents are more reliable than a year ago, though they still make mistakes.
The results were mixed. Asked to spend 30 minutes researching a news pitch, only Muse and Dots used most of that time, and the final pitches did not impress his editors. Gemini Spark produced the strongest angle, but then fell short on a lunch order by adding only a single $5 item for his boss. On the proactivity test, only two agents caught the clash between a protected meeting and a personal reminder buried in email.
For marketers, the more telling finding is the travel test. Every agent suggested the same route through China, which the reporter compared to the Lonely Planet effect. If people hand shopping, dining and trip planning to agents, the same few stores, restaurants and destinations could keep winning. A leaked instruction set mentioned in the piece also points to Muse favoring well-known stores and checking only three sites per request, according to Wired.
The reporter stresses this is not an endorsement. The agents work best with access to email and calendars, which raises privacy and security questions, and his tests cover only a sliver of what they can do. Teams should watch how agents choose sources and merchants, and check how their own brands appear when an agent does the browsing.
Sources
This brief is our own summary. Read the full reporting at the source.
