WIRED finds only two of six AI agents caught a planted scheduling conflict

Giving an AI agent access to email and a calendar did not guarantee it would connect the dots. In a 24-hour comparison published October 27, WIRED’s Reece Rogers found that only OpenAI’s Dots and Anthropic’s Claude Duet warned him about a deliberately planted scheduling conflict. Four other agents missed the overlap before it happened.

Rogers set up six always-on agents on his iPhone: Dots, Claude Duet, Meta’s Muse, Google’s Gemini Spark, xAI’s Grok Bot and Instinct. He designed five everyday tasks to assess their initiative, ability to follow instructions and browser use, scoring each attempt from 1 to 5. He connected his personal Gmail and Google Calendar rather than opening his entire digital life to the tools.

The services came with different entry costs. Rogers listed Dots at $100 a month, Grok Bot at $30, and Gemini Spark and Claude Duet at $20 each. Muse was free, as was invite-only Instinct. Spark was a beta; Duet was a wait-listed research preview. Those were the prices and access conditions described in his comparison.

Morning: the best pitch was not the most patient

For a Thursday editorial meeting, Rogers asked each agent to research technology news and suggest a headline suited to his coverage. He supplied detailed context and asked for 30 minutes of research. Only Muse and Dots kept researching for most of that period; the others answered immediately or after a few minutes.

Following the requested timetable did not produce the strongest idea. Rogers gave Gemini Spark 5/5 for a pitch contrasting retailers’ approaches to shopping agents. Claude Duet earned 4/5, Muse 3/5 and Dots 2/5. Grok Bot and Instinct each received 1/5. The distinction was between obeying a process instruction and delivering an editorial angle the reporter found useful.

Lunch: a plausible restaurant was only the start

The next task added practical constraints: Vietnamese takeout near WIRED’s office, a vegan meal for Rogers’ boss, and meat and caffeine for Rogers. All six agents selected either DragonEats or Freshroll, which he considered reasonable nearby choices. Their carts, however, differed in price and how well they met the brief.

Dots earned 5/5 for the DragonEats order Rogers actually placed, with a $30.57 subtotal before tax and fees. Muse and Grok Bot each scored 4/5. Claude Duet and Instinct built carts he judged too expensive. Gemini Spark fell to 1/5 because its proposed lunch for his boss consisted of a single spring roll. Finding a suitable restaurant had not ensured a suitable meal.

Afternoon: noticing a problem without being asked

Before testing began, Rogers created an immovable 2 p.m. source meeting in Google Calendar. He also emailed himself a fictional scenario: his mother was leaving hospital, and he needed to call her at the same time. The challenge was to read across both services, notice the collision and warn him before it occurred—not simply respond to a question about his schedule.

Dots flagged the clash almost immediately and suggested calling before the meeting or after 3 p.m., earning 5/5. Claude Duet warned him hours ahead but offered no alternative, earning 4/5. After missing the conflict, Muse and Instinct acknowledged that they had not cross-checked the inbox and calendar. Gemini Spark explained that the reminder had arrived by email rather than as a calendar entry or scheduled task.

Rogers cautioned that the exercise covered only a limited slice of the agents’ abilities and was not an endorsement. He also highlighted privacy and security concerns: these assistants work best with personal data, which requires trust. His test left that bargain sharply defined—access to an inbox and calendar was necessary for the planted challenge, but access alone did not make an agent reliably proactive.