Back to the trend of the day: Anthropic's Claude Duet app lets Claude work on its own all day

Poly Rating · from the trend of the day

Claude Duet vs Muse vs Dots: Top 10 AI Agents 2026

Every new all-day agent stops before it sends or pays, but only Claude Duet also refuses some moves outright, such as securities trades and transfers between accounts, even when you say yes, and for now it is a US-only waitlist.

No. 1: Claude Duet
10 places
Current as of October 23, 2026, 14:14 EDT

01Final list

The leading AIs built their own top-ten lists independently. Above is one combined result that is easy to check: every place has a reason, the AIs that included it and, where the AIs gave one, a source.

Order: from most to least AI agreement.

A bright badge means this AI put the entry on its list. A dim badge means it did not.

1
Claude Duet
Strictest written rules: asks before it sends, pays, signs or deletes, and refuses securities trades even with approval. But it is a US Pro/Max waitlist, and METR has no Opus 5.5 figure.
Included it: Claude, ChatGPT, Meta, Grok, Gemini, MiMo, DeepSeek, Kimi.
In their lists: Claude: 1 · ChatGPT: 1 · Meta: 1 · Grok: 1 · Gemini: 1 · MiMo: 3 · DeepSeek: 1 · Kimi: 3
What each AI noted 8
- Claudeplace 1
  It runs on Opus 5.5 and checks your accounts every few hours. It asks before it sends, pays or deletes, and it is the only one of the three new agents with a stated approval rule for deletion.
  Source ↗
- ChatGPTplace 1
  An overnight-capable work partner that explicitly keeps sending, paying, signing, and deleting behind an approval gate.
  Source ↗
- Metaplace 1
  Opus 5.5 inside — Opus 4.6 proxy scored 12.0h 50% horizon (5.3–60.6h CI) and 70min 80% horizon, yet blocks every send/pay/sign/delete behind your tap with a dedicated PII leak classifier.
  Source ↗
- Grokplace 1
  Asks before every send, pay, sign or delete while running on Opus 5.5 with a dedicated leak classifier.
  Source ↗
- Geminiplace 1
  Runs 24/7 on isolated virtual cloud machines to prepare daily briefs, yet enforces strict hard-pauses before sending, paying, or deleting.
- MiMoplace 3
  The only assistant with actions that stay forbidden even after you say yes - no trading securities, no moving money between accounts.
  Source ↗
- DeepSeekplace 1
  The agent that checks your accounts every few hours — and refuses to trade securities even if you say yes.
  Source ↗
- Kimiplace 3
  Three days old and already the most-waitlisted agent — 6,000 connectors, asks before it sends, pays or deletes.
  Source ↗
8 AIs put it here
Source ↗
2
OpenAI Dots (ChatGPT)
Finest controls: per-action Custom Rules, read-only background research, and password changes always left to you. But Pro costs $100/month, excludes EEA, UK and Switzerland, and ChatGPT for Teens is rated Unacceptable.
Included it: Claude, ChatGPT, Meta, Grok, Gemini, MiMo, DeepSeek, Kimi.
In their lists: Claude: 2 · ChatGPT: 3 · Meta: 2 · Grok: 3 · Gemini: 2 · MiMo: 2 · DeepSeek: 2 · Kimi: 2
What each AI noted 8
- Claudeplace 2
  Its Custom Rules let you allow, require approval for, or block each action. Background research is read-only, but the cheapest plan that gets it is $100 a month.
  Source ↗
- ChatGPTplace 3
  More than 4,000 connected apps and a cloud computer turn a conversational assistant into a persistent digital worker.
  Source ↗
- Metaplace 2
  One Dot included in $100/month Pro, each with its own cloud computer and browser across 4000+ apps, but creation desktop-only and Pro excludes EEA, UK, Switzerland.
  Source ↗
- Grokplace 3
  Custom Rules let you set act-without-asking, ask-first or hand-off, with auto-review for account-affecting steps on GPT-6 Astra.
  Source ↗
- Geminiplace 2
  Maintains persistent background memory across desktop and voice interfaces, checking in proactively when cross-app blockers arise.
- MiMoplace 2
  The model that can hold a task for ~16 hours alone - and the only assistant barred from users under 18 from day one.
  Source ↗
- DeepSeekplace 2
  Custom Rules turn 'ask me' into a per-action dial — 4,000+ apps, but no address of its own at launch.
  Source ↗
- Kimiplace 2
  4,000+ app plugins and a Codex hand-off — but at $100/month and no access in the UK, Switzerland or EEA.
  Source ↗
8 AIs put it here
Source ↗
3
Meta Muse
The only free way in, with a separate Sentinel agent that approves outgoing actions. Yet AP reported an agent sharing a user's home address on Marketplace, and Amazon blocked its shopping.
Included it: Claude, ChatGPT, Meta, Grok, Gemini, MiMo, DeepSeek, Kimi.
In their lists: Claude: 6 · ChatGPT: 2 · Meta: 3 · Grok: 2 · Gemini: 4 · MiMo: 1 · DeepSeek: 4 · Kimi: 1
What each AI noted 8
- Claudeplace 6
  It is free at the basic tier and can send email and pay, yet Amazon blocked it 12 days after launch.
  Source ↗
- ChatGPTplace 2
  The broadest consumer ambition in this field comes with a real-world warning: an agent once disclosed its user's home address during a sale.
  Source ↗
- Metaplace 3
  Free tier covers most users, $20 Power (500M tokens/week) and $100 Maximum, launched Sep 8 2026 on iOS/Android/muse.ai/WhatsApp, Secure VM now, Confidential VM promised later 2026.
  Source ↗
- Grokplace 2
  10M+ downloads in US/Canada with free tier, but 'Allow Always' once let an agent share a home address on Marketplace.
  Source ↗
- Geminiplace 4
  Pairs conversational personality with background calendar and shopping tracking, but faces pushback over pervasive Meta-ecosystem data scraping.
- MiMoplace 1
  The only agent where a second AI (Sentinel) must approve everything before it touches the internet - and it still starts free with a one-time virtual card for every purchase.
  Source ↗
- DeepSeekplace 4
  10 million users, a free tier, and a home address shared without a human yes.
  Source ↗
- Kimiplace 1
  The only frontier personal agent with a free tier — and it tops app-store charts while rivals charge $100/month.
8 AIs put it here
Source ↗
4
Gemini Spark (Task mode)
Runs on Google Cloud VMs and asks before spending money or sending email, from $19.99 a month. Common Sense rates Gemini High risk for teens, and it launched US-first.
Included it: Claude, ChatGPT, Meta, Grok, Gemini, MiMo, DeepSeek, Kimi.
In their lists: Claude: see list · ChatGPT: see list · Meta: 6 · Grok: 5 · Gemini: see list · MiMo: 5 · DeepSeek: 3 · Kimi: 4
What each AI noted 8
- Claude
  Named differently in its own list: open the full list.
  see list
- ChatGPT
  Named differently in its own list: open the full list.
  see list
- Metaplace 6
  Gemini 3.1 Pro hits 6.4h 50% horizon (3.9–11.6h CI) and 90min 80% horizon, universal agent for work with its own Workspace email, yet Common Sense rates Gemini High Risk for kids.
- Grokplace 5
  Background agent on Google Cloud VMs that asks before sending, purchasing or form submission; integrated as Task mode.
  Source ↗
- Gemini
  Named differently in its own list: open the full list.
  see list
- MiMoplace 5
  900 million monthly users got a background agent before most of them noticed - and its payment rules are the only ones donated to a public standard.
  Source ↗
- DeepSeekplace 3
  The agent that became a Task mode — and asks for manual confirmation on every write.
- Kimiplace 4
  The first Big-Three always-on agent (May 2026) — but ambient inbox surveillance trades privacy for persistence.
  Source ↗
8 AIs put it here
Source ↗
5
Grok Bot (xAI)
Signs into apps as you and only comes back when it needs approval. But it is still in beta, xAI publishes no hard limits, and Common Sense rates Grok Unacceptable for teens.
Included it: Claude, ChatGPT, Meta, Grok, Gemini, MiMo, DeepSeek, Kimi.
In their lists: Claude: 7 · ChatGPT: 4 · Meta: 5 · Grok: 4 · Gemini: 9 · MiMo: see list · DeepSeek: 5 · Kimi: 5
What each AI noted 8
- Claudeplace 7
  Its background tasks and /loop have a 60-second minimum interval, yet xAI has no consumer 'work while you sleep' announcement we could find.
  Source ↗
- ChatGPTplace 4
  Its own computer and always-on agent teammates make it a serious worker, but a beta is not proof of dependable delegation.
- Metaplace 5
  Each Bot gets persistent cloud PC with browser, filesystem, terminal, works around the clock and is steered only by messaging in Grok app — built for multi-hour rollouts.
- Grokplace 4
  Teams of bots on a shared cloud computer since Aug 2026, with ask-first rules you set and Premium access from $30/mo.
  Source ↗
- Geminiplace 9
  Monitors real-time breaking developments across X overnight, but features looser autonomous conversational filter guardrails compared to peers.
- MiMo
  Named differently in its own list: open the full list.
  see list
- DeepSeekplace 5
  Beta since 11 August; each Bot signs into your accounts and comes back only when it needs a decision.
- Kimiplace 5
  Agents that log in as you and can be taught a workflow by screen recording — with no isolation between your bots.
8 AIs put it here
Source ↗
6
ChatGPT Scheduled Tasks and Agent Mode
Estimate: This is the agent most people already own, but tasks run at most once an hour and pause when idle. In one small, anecdotal test, Agent Mode failed 5 of 10 real tasks.
Included it: Claude, ChatGPT, Meta, Grok, MiMo, DeepSeek, Kimi. Did not include: Gemini.
In their lists: Claude: 3 · ChatGPT: 10 · Meta: 4 · Grok: 7 · MiMo: 7 · DeepSeek: 6 · Kimi: 6
What each AI noted 7
- Claudeplace 3
  It can monitor and ping you only when something changes, but runs at most once an hour and pauses itself if left idle.
  Source ↗
- ChatGPTplace 10
  The easiest entry point to proactive help is also the least convincing substitute for an agent with a persistent computer.
  Source ↗
- Metaplace 4
  Shows every click, search, and page load in real time, pauses for clarification, but help page now says agent mode no longer available and points to Work and cloud browser.
  Source ↗
- Grokplace 7
  Unacceptable for teens per Oct 2026 tests that found zero parent alerts on hour-long crisis chats.
  Source ↗
- MiMoplace 7
  Still the assistant most people actually own: Work mode hands over finished files while its sibling dot does the 24/7 watching.
  Source ↗
- DeepSeekplace 6
  The older agent mode inside ChatGPT — broad reach, but less purpose-built approval control than Dots.
  Source ↗
- Kimiplace 6
  The broadest web action surface of any assistant — as long as the browser tab stays open.
  Source ↗
7 AIs put it here
Source ↗
7
Microsoft Copilot Autopilot (Scout)
Every action runs under a governed company identity that IT can audit, and sensitive steps can require sign-off. But it is still a private preview that needs Frontier enrollment.
Included it: Claude, Meta, Gemini, MiMo, Kimi. Did not include: ChatGPT, Grok, DeepSeek.
In their lists: Claude: 5 · Meta: 8 · Gemini: 3 · MiMo: 4 · Kimi: 10
What each AI noted 5
- Claudeplace 5
  It is the main answer for Windows users, but it is off by default and was limited to Insiders, with no EEA access at first.
  Source ↗
- Metaplace 8
  Agent Mode now default in Word/Excel/PowerPoint, best native app for Windows 11 task bar, three modes Home/Code/Autopilot announced Sep 25 2026, but Incogni puts Copilot among highest privacy risk.
- Geminiplace 3
  Deeply embedded in Windows and Microsoft 365, turning routine inbox chores into automated overnight tasks under strict enterprise tenant policies.
  Source ↗
- MiMoplace 4
  The only agent with its own corporate ID badge: every action runs under a governed Entra identity your IT department can already audit.
  Source ↗
- Kimiplace 10
  Enterprise coworkers with usage-based billing — powerful at work, but no consumer 'act while you sleep' mode.
  Source ↗
5 AIs put it here
Source ↗
8
OpenClaw (self-hosted agent)
The open-source agent that started the always-on wave, and Microsoft's Scout builds on it. You set every permission yourself, and large banks warned in September 2026 of fraud risk.
Included it: ChatGPT, MiMo, Kimi. Did not include: Claude, Meta, Grok, Gemini, DeepSeek.
In their lists: ChatGPT: 8 · MiMo: see list · Kimi: 8
What each AI noted 3
- ChatGPTplace 8
  Maximum configurability is not the same as a safe default: the operator owns the permissions, credentials, and recovery plan.
  Source ↗
- MiMo
  Named differently in its own list: open the full list.
  see list
- Kimiplace 8
  The open-source agent that started the always-on craze — banks now warn it is a fraud target.
  Source ↗
3 AIs put it here
Source ↗
9
Perplexity Comet
A free agentic browser on desktop and phones that asks before sensitive steps, but Common Sense rated Perplexity Unacceptable for kids in September 2026.
Included it: Meta, MiMo. Did not include: Claude, ChatGPT, Grok, Gemini, DeepSeek, Kimi.
In their lists: Meta: 7 · MiMo: 9
What each AI noted 2
- Metaplace 7
  Chromium-based browser free since March 2026 on desktop and mobile, automates booking, email, forms across tabs, but audit flagged covert MCP API with persistent extension access.
  Source ↗
- MiMoplace 9
  The agent that lives inside your browser - and the second assistant Common Sense Media stamped Unacceptable for kids in two months.
  Source ↗
2 AIs put it here
Source ↗
10
Poke
It messages you first inside iMessage, WhatsApp and Telegram rather than waiting to be opened. Only one source covers it, and Cognition bought it in July 2026.
Included it: MiMo. Did not include: Claude, ChatGPT, Meta, Grok, Gemini, DeepSeek, Kimi.
In their lists: MiMo: 10
What each AI noted 1
- MiMoplace 10
  Theonly assistant Apple approved inside iMessage - it texts you first, and one user was reportedly charged 136,000 in a month.
1 AIs put it here
no source given

More about the top three ↓

02More about the top three

The ranking is above. Here is more about the three places at the top: what they are, why they are here, and what is worth knowing.

The three are not measured the same way: Duet is judged on written approval rules, Dots on per-action controls, and Muse on a separate checking agent, so the metrics are not like for like.

1Claude Duet

No. 1 — Claude Duet. On the lists of 8 of 8 AIs; median place there: 1. The AIs put it first for the strictest written rules: it asks before sending, paying, signing or deleting, and refuses securities trades even with your yes.

Did you know

Claude Duet doesn't wait for you to ask. It checks your connected apps every few hours, sends a short morning briefing, and works in its own cloud virtual machine with its own browser, so it never touches your computer.
Source ↗

What it is: A background agent that checks your accounts every few hours, in its own cloud machine.
Its edge: Only one with a stated rule for deletion and outright refusals even when you approve.
Who it suits: Suits careful US Pro or Max users willing to wait; not anyone outside the US.

Trust to act aloneAsks before 4 actions

Why it is here

Four actions always wait for your approval: sending a message, making a payment, signing a document and deleting anything. Duet shows what it is about to do first. Securities trades are refused even if you approve.
Source ↗
Duet can reach more than 6,000 connectors in the Claude Marketplace, per Anthropic's count. Even so, the same four approval rules apply. It checks your accounts only "every few hours", so it does not act nonstop.
Source ↗

Did you know

Duet has no separate price, but its background checks count against your Claude plan limits. An app that keeps watching your email and calendar can use up your allowance even on days you never open it.
Source ↗

History

On October 21, 2026, Anthropic opened Claude Duet as a research preview. Only US Pro and Max subscribers can join, and only from a waitlist. Team and Enterprise plans are left out for now.
Source ↗ ✓ confirmed by several AIs

Keep in mind

In one hands-on test, Duet stalled on some real work, such as inbox, Slack and deck tasks. It also felt slow, doesn't message you first, and has no desktop app yet.
Source ↗

What people say

"But polish isn't enough when you're competing with every other personal agent for attention." — Claire Vo
Source ↗

Where to go further

Open the official Duet page to see how to join the waitlist and what the app does. First read the approval part: Duet shows its planned actions and waits for your yes before it sends, pays, signs or deletes.
Source ↗

Compare the top three ↓

2OpenAI Dots (ChatGPT)

No. 2 — OpenAI Dots (ChatGPT). On the lists of 8 of 8 AIs; median place there: 2. The AIs rank it second for the finest controls: you set each action's rule and passwords stay with you, but it costs $100 a month.

OpenAI · Public domain · Wikimedia Commons

Did you know

When the past-chats memory launched, it skipped the EU, UK, Switzerland, Norway, Iceland and Liechtenstein, reportedly for privacy reasons. It also switches off automatically in Temporary chat mode.
Source ↗

What it is: A persistent ChatGPT worker with its own cloud computer, connected to 4,000+ apps.
Its edge: Custom Rules let you choose act, ask first or hand off for each action.
Who it suits: Suits people who want fine control and can pay $100; not EEA, UK or Swiss users.

Trust to act aloneCustom Rules per action

Wikipedia · Wikidata

Why it is here

Each Custom Rule picks one of 4 behaviors: act without asking, act if pre-approved, ask first, or hand off to you. Changing a password always needs you, whatever your rules say.
Source ↗
When you are not active, a dot's proactive research is limited to read-only app access. Pro excludes the EEA, Switzerland and the UK. CBS reports Pro plans start at $100 a month.
Source ↗

History

ChatGPT launched as a free research preview on November 30, 2022. OpenAI invited people to try it and share feedback, using the public release to learn about the system’s strengths and weaknesses. — OpenAI
Source ↗

Keep in mind

Main complaint: a Plus user reports no warning when memory is full and no usage bar. ChatGPT still confirms new entries as saved, so you can't tell when to clean up.
Source ↗
Users also report ChatGPT saying "Memory updated" while the entry does not show up in the saved-memories list. This matches the ranking's warning that the list may not show everything.
Source ↗

Video: The Man Behind ChatGPT (Sam Altman) · ColdFusion

Compare the top three ↓

3Meta Muse

No. 3 — Meta Muse. On the lists of 8 of 8 AIs; median place there: 2.5. The AIs put it third: it is the only free way in with a separate approval agent, but real mishaps and an Amazon block hurt it.

Did you know

Muse launched September 8 and reached No. 1 on Apple's US free iPhone chart by September 18. Amazon blocked its shopping on September 20, just 12 days after launch.
Source ↗

What it is: Meta's personal agent in the app, with a free tier and paid plans.
Its edge: Only one with a free tier, plus a separate Sentinel agent checking outgoing requests.
Who it suits: Suits people who want to try for free; not those wary of Meta's data or shopping blocks.

Trust to act aloneSeparate Sentinel agent

Wikipedia · Wikidata

Why it is here

Sentinel is a separate agent that reviews every outbound request against your permissions, then allows it, denies it or asks you. Sending emails and making purchases always require your verification.
Source ↗
Muse is the only one with a free tier, up to 100 million tokens a week. The $20 tier gives 500 million and the $100 tier 3 billion. It is US-only, for ages 18 and over.
Source ↗

Did you know

A survey of 1,500 US consumers found only 8% would trust Meta with their passwords. Meta says Muse can't see passwords or card details, and a separate agent called Sentinel approves its internet traffic.
Source ↗

History

In September 2026, Amazon blocked Meta's Muse agent from shopping on its site. Amazon said Muse did not identify itself as an AI agent and seemed to store customers' logins. Users who tried got a popup saying the access broke Amazon's rules.
Source ↗ ✓ confirmed by several AIs

Keep in mind

A Toronto YouTuber let Muse handle his Marketplace messages with "Allow Always" on. Muse reportedly gave a buyer his home address, and the buyer later showed up. Meta says it will clarify the permission prompt.
Source ↗
Amazon cut Muse off from Amazon.com twelve days after launch. It says Meta never warned it, the agent doesn't identify itself, and it seems to store customer logins. Meta says Muse can't see passwords or payment methods.
Source ↗ ✓ confirmed by several AIs

What people say

"I found Muse to be the most useful AI app I had ever used." — Eli Tan, New York Times
Source ↗

Where to go further

Read Meta's launch page for Muse. First look at Sentinel, a separate agent that must approve anything Muse sends to the internet and asks you before sensitive actions like purchases or emails.
Source ↗

MiMo, Kimi put it first. What they saw in it

MiMo: The only agent where a second AI (Sentinel) must approve everything before it touches the internet - and it still starts free with a one-time virtual card for every purchase.
Kimi: The only frontier personal agent with a free tier — and it tops app-store charts while rivals charge $100/month.

Video: Meta Muse Spark Explained: What It Means For Your Business (2026) · Nathan Visser

Compare the top three ↓

The top three side by side

	Claude Duet	OpenAI Dots (ChatGPT)	Meta Muse
In the AIs' lists	8 / 8	8 / 8	8 / 8
Trust to act alone	Asks before 4 actions	Custom Rules per action	Separate Sentinel agent
Put first by	Claude, ChatGPT, Meta, Grok, Gemini, DeepSeek	—	MiMo, Kimi

Places with the same number of AIs are ordered by their average place in the AIs' own lists.

03Worth knowing

What the AIs noticed while building the list.

For 'mostly works', the time horizon is far shorter: Claude Mythos Preview has a 50% horizon of at least 16 hours but only 3h06m at 80%. Opus 4.6 is about 12 hours vs about 1h10m.

Noticed by:

Included it: Claude, Meta, DeepSeek, Kimi. Did not include: ChatGPT, Grok, Gemini, MiMo.

01Final list

Claude Duet

OpenAI Dots (ChatGPT)

Meta Muse

Gemini Spark (Task mode)

Grok Bot (xAI)

ChatGPT Scheduled Tasks and Agent Mode

Microsoft Copilot Autopilot (Scout)

OpenClaw (self-hosted agent)

Perplexity Comet

Poke

02More about the top three

1Claude Duet

Did you know

2OpenAI Dots (ChatGPT)

Did you know

3Meta Muse

Did you know

The top three side by side

More ratings

03Worth knowing

04Questions people ask

05How we counted

1The question

2Why this topic

3Where it applies

4What we looked at

5Who ranked

6How the places were set

06Each AI's own list

What Claude noticed

What ChatGPT noticed

What Meta noticed

What Grok noticed

What Gemini noticed

What MiMo noticed

What DeepSeek noticed

What Kimi noticed