Under Meta's launch video for Muse, the top comment (146 likes) asks a fair question: demos of Muse and its rivals all seem to make dinner reservations and book event tickets, and the commenter doesn't need AI for either on a daily basis. The question gets at the design. A reservation is a low-stakes test of the three things a personal agent has to do: hold context about you, act in another app, and ask before creating a consequence. Below: what the four launches share, why the demos start small, and what to check before you hand one access.
What a personal AI agent is
A personal AI agent is a program that knows things about you, works on tasks without being prompted each step, and acts in your other apps: sending email, filling out forms, making purchases, adjusting plans. Anthropic calls its version a "proactive agent." OpenAI's and xAI's announcements use the phrase "always-on." The label differs, the mechanics are shared:
- It remembers you. Muse "remembers what matters to a person" and can act on details you mentioned once. dots "learn from feedback over time" and pick up your preferences and standards. Grok Bot "remembers conversations" and "gets sharper the more you work together."
- It acts outside the chat. Each agent runs on its own computer in the cloud, with a browser and connections to your apps. Muse can open a browser, fill out forms, and negotiate on your behalf.
- It checks in. Each announcement describes the agent coming back when approval or attention is needed: Muse before it sends an email or makes a purchase, Duet when something needs your attention, Grok Bot "only when something needs your approval."
A chatbot answers when you type. A personal agent advances work while you're away and comes back when something needs you.
The four launches, in order
Four launch announcements define the current snapshot:
| Product | Company | Launched | Where it runs |
|---|---|---|---|
| Grok Bot | xAI | Aug 11, 2026 (beta) | Desktop and iOS app for SuperGrok and Cursor subscribers; agents share a cloud computer |
| Muse | Meta | Sept 8, 2026 | iOS, Android, muse.ai, and WhatsApp; free for most tasks, subscriptions for more |
| dots | OpenAI | Sept 29, 2026 | Inside ChatGPT for Pro, Business Premium, and Enterprise plans in eligible markets |
| Claude Duet | Anthropic | Oct 21, 2026 | Research preview for Claude Pro and Max plans, via waitlist |
Features will move for months, so treat the table as a snapshot from the announcements. If you want to try one, the choice is mostly made for you: it's whichever runs on a subscription you already have.
Why the demos all show a dinner reservation
The reservation example shows up in demo after demo for reasons of demo design.
A reservation is a self-contained job: a visible start, a visible end, minutes of runtime, and small damage if it fails. Video demos need that shape.
It exercises the machinery that matters. Calendar access, form-filling, payment, an approval before money moves: the booking demo walks the same pipes the bigger jobs run through, on stakes small enough to demo in public.
And it's universal. A booking demo is easy to follow in fifteen seconds; an invoice-reconciliation demo asks its audience to care about a workflow only some of them have.
Meta's own launch post leans domestic: Muse turns a saved recipe reel into a grocery list, suggests a dinner-party menu, and remembers guests' dietary restrictions before sending invites. The other announcements go further into money and errands. Muse's examples include selling a car for more, lowering a bill, and adjusting a training plan as the rest of someone's life shifts. OpenAI describes an early tester whose dot noticed he'd forgotten to invoice a publication, prepared the invoice, and sent it after his approval. Anthropic positions Duet for "professionals and small business owners with more on their plates than they can track."
The commenter hopes these agents aim higher than reservations. They do, on a ramp: the small example is the trust scaffold, a low-stakes job done with full visibility, and the bigger job comes after the approval behavior earns it. The ceiling on these products is the access you grant, not the demo. Muse can negotiate on your behalf and pay with a one-time-use card, but only inside the connections and permissions you set.
The job underneath the demos
A comment on the Muse video describes the appeal in one line: the commenter isn't a power user, just busy and scattered, and wants one chat for stray thoughts and reminders, plus an answer to "what did I forget?" The four announcements all build toward that. Duet sends a morning brief across your connected apps, then messages you only when something needs your attention. Muse keeps working after you close the app and comes back when something changes. dots watch connected apps in the background and, in OpenAI's example, start investigating the moment a bug shows up in Slack. Grok Bot works through multi-step jobs around the clock and reports back when it's finished. The shared promise: the agent holds a thread, advances it, and brings you back in when judgment or approval is needed.
Before you hand over access
The access that makes these agents useful is the same access that makes them a decision worth making deliberately. An agent that can send email and spend money can also send the wrong email and spend money badly, so the announcements describe different limits and approval controls.
- Start read-only. dots' background work runs on tools "restricted to be read-only," which means they can't send messages, change app content, or control your browser or computer. Muse lets you set whether an email connection can read only or also send on your behalf.
- Keep approval on for consequences. Muse checks before it sends email or makes a purchase, and it shows "a complete audit trail" of its actions and its planned ones. Anthropic says Duet checks with you before it acts. dots ship with built-in rules for when to act alone versus ask, and Custom Rules let you require approval for specific actions.
- Know where the credentials sit. dots can use saved passwords without exposing them to the model, and Muse stores credentials where the agent can use them without seeing them. Grok Bot signs into the tools you already use, so it's worth reading what it holds before connecting anything sensitive.
- Check the data settings. Muse keeps VM data out of Meta's ad systems and offers a training opt-out, and Meta says a Confidential VM arrives later this year, encrypted with a key only the account holder holds. OpenAI doesn't use Business, Enterprise, or Edu workspace content to train on by default.
We'd start with one connected app and one small job, watch the approval prompts for a week, then decide what to add. The proportionality rule from our verification post holds here: low-stakes output can ride on judgment, high-stakes output gets checked, and an agent with purchase power is high stakes. For what happens to what agents see, Do AI Agents Train on Your Data? What Agents See, Save, and Learn walks through the vendor policies, screenshot retention, and prompt injection.
FAQ
What is a personal AI agent?
An assistant that holds context about you, works in the background without being prompted each step, and acts in your other apps within the permissions you set. Four launched between August and October 2026: xAI's Grok Bot, Meta's Muse, OpenAI's dots, and Anthropic's Claude Duet.
How much does Muse cost?
Free for most of what people need, with subscription plans for people who want to do more, per Meta's announcement. CNBC reported in October 2026 that it's free for now while Meta works out the business model around it.
Which personal AI agent should I try first?
The one attached to a subscription you already pay for, with one app connected and one small job: dots run on ChatGPT Pro, Business Premium, and Enterprise plans; Duet is a research preview for Claude Pro and Max behind a waitlist; Muse is free to start in the US; Grok Bot is in beta for SuperGrok and Cursor subscribers.
Are personal AI agents safe to connect to email?
Email is a reasonable first connection, in read-only mode. Turn on approval prompts before granting send-on-your-behalf access, and keep the audit trail on where the product offers one: Muse shows its actions and its planned ones. If the agent starts acting without asking, revoke the send access first.
Can a personal AI agent spend money?
Yes, with approval in front of it. Muse checks before any purchase and pays with Link's one-time-use card, so your card details stay hidden, and it's the first agent covered by Link's purchase protections. On dots, you can set Custom Rules that require approval for specific actions like spending.
Want to go further?
If you try one, we'd love to hear how the first week of approval prompts goes. The posts below cover the layers this category stands on:
- What Is an AI Agent? Explained Simply: the definition this category builds on.
- What Are AI Agents in 2026?: the wider agent landscape, including prompt injection and the questions that sort the label.
- How to Verify AI Output: A Simple Framework: when to trust and when to check, the core skill behind agent permissions.
- What Is MCP? Model Context Protocol Explained Simply: how agents connect to tools in the first place.
- DIY AI Agent or All-in-One Bot? How to Decide: for readers already running a DIY setup and weighing a switch.