AI Next WaveFour Agents Want Your Inbox, Your Calendar and Your Credit Card. Only One Can Show Its Work.
Anthropic launched Claude Duet on Wednesday, the fourth proactive agent in six weeks asking for your apps, your logins and your purchasing power. Duet, OpenAI's Dots, Meta's Muse and xAI's Grok Bot all promise to work while you sleep. They differ sharply on who holds your credentials, what stops them from spending, what their background work costs you, and whether anybody actually uses them.
The new arrival
Anthropic opened Claude Duet on October 21 as a waitlisted research preview for US subscribers on its Pro and Max plans, on iOS, Android and claude.ai. Duet runs on Opus 5.5 and it is built for professionals and small-business owners. It sends a morning briefing drawn from your connected apps, checks those apps every few hours, and works across more than 6,000 connectors, Gmail and Slack among them. It can buy things for your business through Stripe Link and Shop Pay. Before it sends a message, makes a payment, signs a document or deletes anything, it stops and asks you.
Anthropic is the fourth company to ship this product in a single season. Meta launched Muse on September 8, OpenAI launched Dots at DevDay on September 29, xAI's Grok Bot has been in beta since August 11, and Google's Gemini Spark and Microsoft's Autopilot are chasing the same idea. Every major AI company now wants to sell you an agent that acts first and reports back later. So the comparison that matters is what each one does with your access once you give it.
Who acts, and who only watches
Dots is the most cautious design of the four. Each dot gets its own cloud computer and browser, and you give it standing assignments tied to goals. When a dot is idle, its proactive research mode is read-only. It can look through your connected apps for ways to help, and it cannot send a message or change anything in that mode. It reaches more than 4,000 apps through plugins.
Duet and Muse act more freely, with approval gates at the sensitive steps. Muse takes goals and works them end to end, booking, buying and filling out forms, then checks with you before payments and emails go out. Grok Bot sits at the far end. Its Bots are named, persistent teammates with skills and routines. Show a Bot a job once and it learns the workflow by watching, then reruns it on a schedule, which is genuinely useful for a small shop with repetitive work. It also means that Bot lives on the same machine as everything else you connected, which brings us to the part of this comparison I care about most.
Who holds your logins
Duet never sees your passwords, and Anthropic runs a dedicated classifier over its work to guard against the leak that matters in this category: a prompt injection, where instructions hidden in an email or a web page try to talk the agent into handing over something private. Muse uses the strongest separation on paper. Your agent lives in a dedicated Secure VM, keeps your logins in a credential store it can use but cannot read, and a second agent called Sentinel approves anything before it reaches the internet. At checkout, Stripe Link generates a one-time card number, so your real card stays hidden from the merchant and from Muse. Meta also says the data inside that VM is not shared with its ad systems. I note that commitment because Meta wrote it down, and I will hold them to it.
Grok Bot is where I get off. Every Bot on your account shares one cloud computer: the same files, browser sessions and logins. xAI's own documentation says the computer belongs to your account, not to any one Bot, and tells you flatly not to put a credential or file on it if another Bot of yours should not be able to use it. Its words for the separate Bot screens: "not separate security boundaries." Sign one Bot into a tool and that signed-in session is available to every Bot you run. Credit to xAI for printing the warning in its own docs. The warning is still the warning.
What it costs while you sleep
Background work is not free, and the pricing pages are where the promises get quiet. Duet carries no separate sticker price, but every background check counts against the usage limits on the plan you already pay for, so a heavy user hits the cap sooner and finds out mid-task. Anthropic concedes a second catch in its own launch material: Duet can misjudge what is urgent, and it is weaker at personal tasks. Dots avoids the meter for its first month, then counts usage, and OpenAI has not published post-intro pricing. Access is restricted to the expensive seats: ChatGPT Pro at $100, $200 or $500 a month, or Business Premium at $100 per user per month billed annually. Muse is the cheapest door in: a free tier that still requires a payment card, then $20 and $100 a month, with a small business version launched September 29. Grok Bot has no standalone subscription. Per xAI's documentation, it comes included with every paid Cursor plan, individual and Teams alike, or through a linked SuperGrok subscription, and usage resets weekly.
The only scoreboard that counts
Three of these four products have launch events, waitlists and demos. One has a usage record, counted by an outside estimator. Sensor Tower put Muse at 10 million US downloads in its first 40 days and 2.5 million daily users by October 16, with 30% of users still active at day 7 and 23% at day 14 (reported by TechCrunch). Those retention numbers are decent, not spectacular, and they are the only realized adoption figures anyone has for this category. Meta has published no usage figures of its own. Duet is a waitlist, Dots has no public usage numbers, and Grok Bot is a beta bundled into other subscriptions. When a company won't tell you how many people kept using the product after the first week, assume the answer is the reason.
Here is where I land. If you run a small business and the inbox, invoicing and scheduling grind is eating your week, this category is aimed at you, and the hours are real if the safeguards hold. Start with the product whose limits you can read in writing: approval before anything sends, signs, buys or deletes, credentials the agent itself cannot see, and background work on a meter you understand. On the evidence available so far, Muse has the adoption and a serious security design, Duet has the clearest approval gates and the most honest list of its own weaknesses, Dots has the safest idle mode and the steepest price of admission, and Grok Bot has the one architecture I would not connect to anything that matters until the shared computer goes away. Hand any of them one job and one tool first. An agent that earns more access is fine. An agent that starts with all of it is a liability with a friendly name.
References
Reuters, Anthropic unveils Claude Duet AI agent app, October 21, 2026
Reuters, OpenAI takes on Meta with always-on Dots agent in enterprise AI push, September 29, 2026
TechCrunch, OpenAI launches Dots, its agentic avatar, September 29, 2026
TechCrunch, Meta's Muse hits 10M US downloads in 40 days (Sensor Tower data), October 20, 2026
Meta, Muse product page: Secure VM, Sentinel approvals and credential store
xAI, Grok Bot documentation: overview, access and the shared cloud computer
xAI, Grok Bot documentation: the shared computer, shared sessions and credentials