The Best Proactive AI Agents in 2026: Dots, Duet, Muse, Grok Bot, Manus, Instinct and Ari

Seven personal AI agents now promise to work for you around the clock: ChatGPT dots, Claude Duet, Meta's Muse, Grok Bot, Manus's Cue, Instinct and Ari. Here's how they compare on the four things that decide whether you should hand one the keys: privacy, reach, apps and teams.

The short version

As of October 28, 2026. The sources for every cell are in the sections below.

The 3 a.m. test

The proactive AI war is heating up. In ten weeks, six companies shipped an agent that doesn't wait to be asked. It wants your inbox, your calendar, your contacts and, in most cases, your card, and it keeps working while you sleep.

We've been building Ari for the last year and a half, so I'm not surprised the industry ended up here. An assistant that waits for you to type was never going to be the end state. What surprises me is how the race is being run: almost entirely on capability, and barely at all on trust.

That's the part I can't stop thinking about. Before Ariso, I spent years building JupiterOne, a cybersecurity platform, and before that I built systems at Fidelity, IBM and Cisco. In security, you learn to judge a system by what it does when nobody's watching. That's when things go wrong, and that's when you find out whose side it's on.

With these agents, that's every night. At 3 a.m., while you sleep, something is reading your email and acting in your name. The question is who it really works for.

Six agents in ten weeks

• Grok Bot (August 11). From SpaceX's AI division, the company formerly known as xAI, built on Cursor. xAI calls it "your team of always-on agents", each with its own computer. It's included with the Premium, Super and Ultra tiers of The X Subscription, and not with Free or Plus.
• Instinct (late August). A startup's personal assistant you text or call. "It can use a phone and computer in the same way that you do." It's invite-only and free during the beta.
• Muse (September 8). Meta's personal agent, running in what Meta calls a dedicated, secure computer. It's free to start, with paid plans at $20 and $100 a month.
• Manus 2.0 and Cue (September 28). Cue is Manus's new personal-agent app, invite-only for now, where "each agent has its own email, phone number, wallet, and computer." Meta bought Manus last December; after Beijing objected, the deal is being unwound, and Manus says it will return to operating as an independent company.
• ChatGPT dots (September 29). OpenAI's always-on agents inside ChatGPT, each with "their own cloud computer, their own browser, and the apps you've connected". They're for Pro and Business Premium subscribers, and the first dot is included.
• Claude Duet (October 21). Anthropic's proactive agent, in research preview with a waitlist for Pro and Max subscribers in the US. It sends a brief each morning, checks your connected apps every few hours, and asks before it sends a message, makes a payment, signs a document or deletes a file.

These are serious products from serious teams. Most of them can do things Ari can't, and I'll get to those.

But under the hood, most of them are the same idea: a frontier model, a cloud computer with a browser, and a chat. The model isn't where they really differ. It's four other places.

Four questions before you hand over the keys

1. Privacy. Who can read it, and what does it learn from? Not what the settings page offers, but what's on by default.
2. Reach. Does it come to you, or do you have to go to it?
3. Apps. Is there more than a chat?
4. Teams and work. Does it work with the people in your life?

1. Privacy: read the defaults

The first thing you learn in security is that almost nobody changes a default. So I don't care much what a privacy page lets you turn off. I care what's on when you open the app.

Muse. Meta's own help page says the setting that lets your interactions train its models "is on when you first use Muse." Deleting something doesn't mean it's forgotten: "Muse may still remember information it learned from what you deleted." And TIME reported that "each hour, Muse updates its dossiers on you and the people you've mentioned" and that "even people who don't use Muse are subject to this mapping process by other people's Muse agents." To be fair to Meta, it says Muse doesn't share your conversations with its ad systems, and it's building a Confidential VM designed to keep your information private "even from Meta." That isn't available yet.

Instinct. Its terms say it may use what you give it for many things, "including training AI models." You can opt out, but "on a go-forward basis," and "we may still use your Materials for AI model training when your Materials are flagged for safety review." Early testers found the original terms asked for a "perpetual and irrevocable" license. Those terms have since been revised.

ChatGPT dots. On Business and Enterprise, OpenAI doesn't train on your data by default. On personal plans like Pro, ChatGPT's "Improve the model for everyone" setting decides, and it covers your dot's conversations and work. One detail stood out to me: according to OpenAI's dots privacy FAQ, as summarized by Flavio Copes, "you can't view, correct or delete individual dot memories. Deleting the dot is the only way to clear them."

Claude Duet. Duet runs on your Claude Pro or Max account, and on Anthropic's consumer plans your chats are used to improve its models if "you choose to allow us" in your privacy settings. Anthropic says Duet doesn't see your passwords or logins when you give it access to an account, and checks each message with a dedicated safety classifier.

Manus. Its privacy policy says de-identified data may be used to "develop and improve our models." Opting out of training is listed as a Team-plan feature. The policy also shares identifiers with marketing vendors for advertising, which you can turn off.

Grok Bot. Training "follows the applicable Cursor account and privacy settings." And one line from its own FAQ deserves more attention than it gets: "Do not use separate Bots as a security boundary."

None of these companies are villains, and some of these choices are reasonable trade-offs. But notice the pattern. The privacy you get is mostly decided by two things you never see in a demo: the default, and the business model.

The most important privacy feature isn't on the settings page. It's how the company makes money.

Meta's business is ads. Manus's policy shares identifiers with ad vendors. We make money from subscriptions, not ads.

Here's what that lets us promise, and what's on from the first message:

• Ari never trains on your data. If you want a model trained on you, the Pro plan lets you train your own. You pick the sources, review every example, and press start.
• Every person's data is encrypted with a key unique to them. Company data gets its own organization key, and keys are managed in HashiCorp Vault.
• Our own staff can't read it. That includes our database administrators, unless you authorize it for support.
• Your employer doesn't get your conversations. What a company sees from Ari is anonymized and aggregated.
• You can see what Ari knows. Ari's Brain shows it, lets you correct it, and exports all of it.
• You can go off the record. Ask Ari to "talk off the record" for a private session that's forgotten when it ends.
• Closing your account deletes your key. Whatever is left becomes unreadable.

We're SOC 2 Type II and Google CASA certified. The details are on our trust and security page.

2. Reach: does it come to you?

Every one of these companies agrees an agent should live where you are. They just disagree about where you are.

• Dots lives in ChatGPT, on the web, desktop and mobile, and in Slack and Microsoft Teams.
• Duet has its own app on iOS, Android and the web, separate from Claude.
• Muse lives in its own app and in WhatsApp, with Meta's glasses announced next.
• Grok Bot lives in its desktop and mobile apps, on X and in Slack, with voice.
• Manus's Cue works in iMessage, Telegram, LINE and Slack.
• Instinct works by text, WhatsApp and its own email address, and it's bringing itself into group chats.

Notice that the biggest platforms each pull you toward their own: Muse toward Meta's apps, Grok Bot toward X, dots toward ChatGPT, Duet into a new app of its own. That's not a criticism. It's what platforms do.

Ari doesn't have a platform to pull you into, so it goes where you already are:

• Texting: iMessage and text, including group texts. Friends without Ari can be in the thread.
• Work apps: Slack and Teams.
• Email.
• The web.
• Voice.
• Your desktop: Oats, our free notetaker, takes notes on your calls without a bot joining them.

The mobile app and WhatsApp are coming soon. Until then, Ari's app runs in your phone's browser.

What matters more than the list is that it's the same Ari, with the same memory, wherever you reach it. Tell it something in a text and it knows in Slack.

Instinct and Manus are strong here too, and I'd call reach the new frontier of this category. The days of a personal AI you have to go visit are over.

3. Apps: is there more than a chat?

Ask the other six how they get things done and you mostly get the same answer: give the model its own computer and a browser, and let it click around as you. It's a powerful idea. Dots, Duet, Grok Bot, Manus and Muse all work that way, and Instinct "can use a phone and computer in the same way that you do."

It's also why this fall's security stories look the way they do.

• Manus: researchers at Salt Labs showed that a single email could hijack a Manus agent and reach the user's connected accounts. The fix came through Meta's bug bounty.
• Grok Bot: a tech CEO said his Grok Bot posted his bank balances to his company Slack after two of his bots, sharing one computer, "collaborated." That's his account, as Techdirt reported it.

When an agent can do anything a person at a keyboard can do, it can also do anything a person at a keyboard can do.

Duet is the one heading somewhere different. Every's review says it splits recurring topics into their own channels, keeps a running list of what it's working on, and can keep docs, slides and dashboards current. That's closer to apps than to a chat, and I think it's the right instinct.

Ari can work that way too: in your browser, through Ari's extension, when a job needs you signed in (it plans a night out and books it for you to pay, and never handles your passwords), or on a computer of its own in the cloud. But most of what Ari does lives somewhere you can see: in apps.

• Meetings: notes, prep before every call, and what everyone promised.
• Todos and reminders: made from your meetings and mail, closed when they're done, and reminders for other people too ("remind Maya to bring the cupcakes Monday").
• Contacts: they build themselves. Each week Ari picks a few people worth getting back in touch with, and says why now.
• Research: search everything Ari remembers, with its source, and research reports cited to your Google Drive.
• Booking and planning: find a time with anyone, and share your own booking page.
• And more: Projects, your Profile (the card that tells people how to work with you), Personality and Strengths, Circles and Groups, and Routines.

These aren't features hidden behind a prompt. Each one is a first-class use case with its own data, which means you can see it. You can open the todo, fix the contact, delete the note.

A chat is an interface. An app is a promise to keep your data in a shape you can see.

An agent with a private memory you can't view, correct or delete is asking for a lot of trust. Apps are how you earn it.

4. Teams and work: does it work with the people in your life?

This is where the field splits.

• Dots, Grok Bot and Manus are built for work. They sell admin controls, SSO and team plans.
• Duet is built, in Anthropic's words, "for professionals and small business owners with too much to track", but for now it's only on the individual Pro and Max plans.
• Muse has a small-business version that connects to Slack, Asana, QuickBooks and more.
• Instinct is, by its own terms, "for your personal use only."

Look closely at those team plans, though. They buy everyone an agent, on one bill, with an admin panel, and each agent still works for one person. A dot "responds to you" unless you tell it to engage with others. Grok Bot's bots hand work to each other within your account. Cue's agents hand work to each other in a group chat that you direct. None of them describes your agent working with a colleague's: asking Joe's when he's free, having it remind him about Friday, handing him a task, or telling you how he likes to get feedback.

That's the part I think matters most. Work isn't a workspace with an admin panel. Work is people. So is life. The hard part has never been doing a task. It's doing it with someone else.

So every Ari works with everyone else's:

• Ask for Joe's time. "Find 30 minutes with Joe next week." Your Ari sees the availability Joe shares and settles it with his.
• Remind Joe, not just yourself. "Remind Joe to send the deck Friday," and Joe's Ari reminds him.
• Hand Joe a task. Ari tells you when a todo is better off with someone else. Send it, and it lands on Joe's list with your name on it.
• Ask what Joe's like. His Profile tells people how to work with him, not his job history. If he turns it on, the people he chooses can ask his Ari how he works and what he's working on, from only the sources he picks.
• Ari can sit in the group text with the people in your life, and decides when it's actually useful to chime in.
• Feedback about a colleague stays anonymous unless you ask for it to be attributed. Your conversations never go to your employer.

On a team, everyone has Ari, so this works from day one. With friends and family who use Ari, it works the same way.

For teams, the same Ari also runs the whole system: projects that keep their own status, the weekly rhythm, goals and priorities, and how everyone's doing. It works in Slack and Teams. For large companies, we deploy into their own cloud.

And most of us don't have a work life and a personal life anymore. We have one life, with work in it. One Ari across both, with a privacy line between them, is the design.

Where they beat us

No agent wins everywhere. Here's where the others are ahead today.

• Doing errands end to end. Instinct's founder says the platform is approaching a billion dollars in annual transactions, more than half of it travel. Muse, dots and Grok Bot can all make purchases for you too. Ari plans and books in your browser for you to pay. It's not a shopping concierge, and if that's what you want, we're not the best pick.
• Long-running technical work. Dots, Duet, Grok Bot and Manus give every agent a persistent cloud machine that keeps working when your laptop is closed. For long jobs and code, they're further along.
• Making things. Duet can build a web app or keep a deck current, and Every called it "the first personal agent I'd call a pro tool." Ari writes documents to your Drive and builds dashboards you describe, but it can't edit a doc in place yet.
• Phone calls. Instinct is rolling out calls to businesses, and Cue gives agents their own phone numbers. We don't make calls for you.
• A native mobile app. Muse, dots, Duet and Grok Bot have one. Ours is coming.
• Frontier models. OpenAI, Anthropic, SpaceX and Meta train their own. We don't. Ari uses a mix of models and switches as they improve, and you can bring your own OpenAI or Anthropic keys if you'd rather.

So which should you use?

• You live in ChatGPT and already pay for Pro: dots.
• You pay for Claude Pro or Max and your work is documents, decks and code: Duet, once you're off the waitlist.
• You want an agent for long-running, technical work, like code: Grok Bot, if you're in Cursor, or dots.
• You want errands, travel and purchases done for you: Instinct or Muse. Just read the defaults first.
• Your life runs on Telegram or LINE: Manus's Cue, once you get an invite.
• You want a chief of staff for your work and your life, private by design, that works with your colleagues' Aris: that's Ari.

If you're moving over from ChatGPT or Claude, Ari can import the memory you've built there, so you don't start from zero.

Back to 3 a.m.

These agents will keep getting more capable. That part's inevitable, because most of them run on the same handful of frontier models.

What won't converge is what each company does with everything its agent learns about you, and about the people around you, while you sleep. That isn't decided by benchmarks. It's decided by defaults and by business models.

So pick the agent whose defaults you'd have chosen anyway. That's the one you'll still trust at 3 a.m.

---

I'm Erkang, CEO of Ariso. We've spent a year and a half building Ari to earn that trust. Everything here about the other six is linked to its source and current as of October 28, 2026. These products change every week, so if something's out of date, tell me and I'll update it. Ari is free to start, no card needed.