# Testing 6 Always-On Personal AI Agents: How OpenAI Dots, Meta Muse, and Claude Duet Compare

Over the past month, every major AI lab has rolled out an "always-on" personal AI agent for smartphones—software that doesn't just wait for you to type a question in a chat box, but proactively watches your calendar, scans your Gmail inbox, controls its own web browser to build takeout orders, and texts you first when it spots a scheduling conflict. A 24-hour head-to-head stress test of the six biggest personal AI agents on iPhone shows that these digital assistants finally work on real daily chores—and reveals clear winners depending on whether you want to spend $0, $20, or $100 a month.

A year ago, consumer AI agents routinely got stuck on basic login screens or hallucinated flight prices. Today, six competing personal agents are vying for a permanent spot on your phone's home screen: **OpenAI's Dots** ($100/month inside ChatGPT), **Meta's Muse** (Free standalone app), **Anthropic's Claude Duet** ($20/month wait-listed preview app), **Google's Gemini Spark** ($20/month beta inside Gemini), **xAI's Grok Bot** ($30/month standalone app), and **Instinct** (Free invite-only SMS text agent).

To find out which ones actually save time in everyday life—and which ones waste your money or annoy you—software journalist Reece Rogers connected all six agents to his personal Gmail and Google Calendar on a single iPhone and put them through five real-world tests over 24 hours: doing 30 minutes of deep morning work research, building a two-person DoorDash lunch cart with dietary restrictions, proactively catching a 2:00 PM calendar-vs-email conflict before it happened, planning a 2-week vacation itinerary with a strict 4-hotel limit, and handling a late-night "I can't sleep, help me organize my tasks" check-in.

The results were eye-opening: **OpenAI's Dots** took #1 Best Overall for actually following complex instructions, **Meta's free Muse** came in a close second as the easiest and best free option for everyday users, and **Anthropic's Claude Duet** won Best Interface—though users need to watch out for proactive agents quietly burning through their monthly usage limits overnight.

## Fast Facts: 6 Personal AI Agents Tested Over 24 Hours on iPhone

- **Test Benchmark & Source**: 24-hour real-world evaluation of 6 always-on smartphone agents across 5 daily tasks, conducted and detailed in [WIRED's hands-on comparison of personal AI agents](https://www.wired.com/story/i-tested-every-major-ai-agent-for-a-day-heres-the-best-one/).
- **The 6 Agents & Monthly Prices**:
- **OpenAI Dots**: $100/month (built inside ChatGPT + iPhone widget) — **Winner: Best Overall**.
- **Meta Muse**: Free (dedicated smartphone app, topping download charts since September) — **Winner: Most Accessible / Best Free Pick**.
- **Anthropic Claude Duet**: $20/month (dedicated app, wait-listed research preview) — **Winner: Best User Interface**.
- **Google Gemini Spark**: $20/month (beta inside the Gemini app) — **Best news synthesis angle, but failed the takeout cart test**.
- **xAI Grok Bot**: $30/month (dedicated app).
- **Instinct**: Free with invite code (operates purely via SMS text messages) — **Voted Most Annoying**.
- **Proactive Conflict Test Winner**: Only **2 of the 6 agents** successfully cross-checked an incoming Gmail message ("call Mom at 2 PM when she's discharged from the hospital") against an existing 2:00–3:00 PM Google Calendar meeting and texted the user *before* the 2 PM collision.
- **Instruction-Following Test**: When asked to spend a full 30 minutes researching before pitching a work idea, only **OpenAI Dots** and **Meta Muse** actually researched for the requested window; the other four blurted out shallow answers within 2 minutes.

## Buyer's Guide: How All 6 Personal AI Agents Compare on Price & Daily Tasks

The table below compares the six tested smartphone AI agents across pricing, app design, real-world test performance, and their biggest practical drawbacks.

| Personal AI Agent | Monthly Cost | How You Use It | Real-World Test Highlights | Biggest Practical Catch |
|---|---|---|---|---|
| **OpenAI Dots** (#1 Best Overall) | **$100 / mo** | Inside ChatGPT app + iPhone home widget | Followed the 30-min research rule; nailed the $30.57–$43.40 vegan/meat Vietnamese lunch cart & late-night triage | Expensive $100/mo paywall; lives inside ChatGPT rather than a standalone app |
| **Meta Muse** (#2 Best Free / Accessible) | **Free ($0)** | Dedicated mobile app | Easiest setup; followed 30-min research rule; close runner-up across shopping, food, and travel tests | Meta privacy track record; leaked instructions show it checks only 3 mainstream websites per shopping query |
| **Anthropic Claude Duet** (Best Interface) | **$20 / mo** | Dedicated mobile app (Waitlist preview) | Cleanest, easiest app UI for tracking ongoing background tasks; solid 4-hotel China travel route | Proactive background check-ins and overnight tasks **count against your normal Claude chat usage limit** |
| **Google Gemini Spark** | **$20 / mo** | Beta inside the Google Gemini app | Wrote the sharpest morning news briefing contrasting Walmart vs. Amazon AI shopping rules | **Failed the lunch order test** (only put a single $5 spring roll in the cart for the user's boss) |
| **xAI Grok Bot** | **$30 / mo** | Dedicated mobile app | Handled the 13-night Beijing–Xi'an–Chengdu–Shanghai bullet-train itinerary constraint cleanly | Mid-pack proactivity; $10/mo pricier than Claude Duet and Gemini Spark |
| **Instinct** (Most Annoying) | **Free** (Invite code) | Purely via SMS text messages (No app) | Zero app install required; works inside your phone's default text message thread | No visual dashboard to see what its browser is doing; missed the 2 PM email/calendar clash and sent a grimace emoji |

## What Works—And What Still Breaks—When You Let an AI Agent Run Your Day

Seeing how all six agents handled the same five everyday errands shows exactly where personal AI agents can save you an hour today—and where you still need to double-check their work:

1. **Multi-Person Takeout & Errand Carts Work Remarkably Well (With One Exception)**: When asked to build a pickup lunch order near the office for two people—one wanting a meaty dish plus Vietnamese iced coffee, and a vegan boss craving Vietnamese tofu rolls (specifically swapping out fish sauce for vegan soy or peanut sauce)—five of the six agents controlled a virtual browser, found nearby spots (*DragonEats* and *Freshroll*), customized the vegan sauces properly, and pre-built carts ranging from **$22.55 to $53.73**. Only **Gemini Spark** fumbled by adding a single $5 roll for the boss and forgetting the main meal.
2. **Most Agents Still Forget to Cross-Check Email Against Your Calendar**: The biggest promise of an "always-on" agent is that it watches your back without being asked. Yet when a reminder to call Mom at 2:00 PM arrived in **Gmail** while an unmovable meeting was sitting at 2:00 PM on **Google Calendar**, four of the six agents missed the overlap completely—apologizing afterward that they only checked the calendar and forgot to cross-reference unread emails unless told to do so.
3. **The "Lonely Planet Effect" in Travel Planning**: When asked to plan a 2-week trip to China mixing history, food, and nightlife with a strict cap of **4 hotels maximum**, every single agent obeyed the 4-hotel rule—and every single agent picked the exact same 4 cities in the exact same order (Beijing $\\rightarrow$ Xi'an $\\rightarrow$ Chengdu $\\rightarrow$ Shanghai via high-speed rail). Agents are fantastic at logistics constraints ("max 4 hotel moves, no overnight trains"), but they will steer everyone toward the exact same tourist hotspots.
4. **Watch Out for "Proactive" Token Drain and Over-Sharing**: Two practical warnings stood out during the test. First, whenever a proactive agent like **Claude Duet** wakes up on its own to check your inbox, draft a reply you didn't ask for, or build a summary overnight, that background work eats into the same $20/month token allowance you use for regular Claude chats. Second, never connect a personal agent to both your personal bank/finance apps and your work Slack—as one X user learned the hard way when his AI agent cheerfully posted his personal bank account balance into a public company Slack channel.

## Practical Setup Guide: How to Try Personal AI Agents Safely

If you want to start using a personal AI agent on your phone this week without getting hit by surprise bills or awkward privacy leaks, follow these four steps.

### The 4-Step Enterprise Implementation Playbook

1. **Start With Free Meta Muse or $20 Claude Duet Before Paying $100 for OpenAI Dots**: Unless you are a heavy power user who already pays $100/month for OpenAI's top tier, start with **Meta Muse** (Free) to test browser errands like takeout carts and flight comparisons, or join the preview for **Anthropic Claude Duet** ($20/month) if you want the cleanest app interface for organizing daily work.
2. **Use a Dedicated "Agent" Calendar & Email Filter—Never Connect Work Slack**: Never connect a consumer AI agent simultaneously to private personal accounts (banking, health, personal texts) and public team channels (company Slack, Teams, or shared work docs). Connect it only to the specific Google Calendar and email label it needs to read.
3. **Explicitly Schedule a Daily 8:00 AM "Inbox + Calendar" Cross-Check**: Because four out of six agents failed to compare Gmail reminders against Google Calendar events automatically, send your agent a one-time setup rule: *"Every morning at 8:00 AM, cross-check my unread emails and self-sent notes against my calendar for the next 48 hours and text me if any times conflict."*
4. **Turn Off Unsolicited Overnight Tasks to Save Your Monthly Message Cap**: On $20/month plans like Claude Duet and Gemini Spark, configure the agent's proactivity settings so it only runs background browser sessions during set hours (e.g., 8 AM and 5 PM) rather than burning through your monthly chat quota while you sleep.

## Next Steps

- **The "Build My Cart, Don't Checkout" Test**: Ask Meta Muse or OpenAI Dots to build a customized lunch or grocery pickup cart for two people with specific dietary swaps, and have it stop at the final review screen so you can verify the items and prices ($20–$45) before tapping Pay yourself.
- **The Constraint-Based Trip Draft**: Give an agent a messy logistics puzzle—such as *"Plan a 10-day trip across Italy or Japan with no more than 3 hotel changes and under 2 hours of train travel between stops"*—then verify the local restaurants yourself on YouTube or Google Maps.
- **Audit Your Connected App Permissions Monthly**: Once a month, open your Google Account (`myaccount.google.com/permissions`) and revoke access for any AI agent apps you tested and stopped using so they don't keep reading your inbox in the background.