dots vs Grok Bot vs Muse vs Gemini Spark: 4 Always-On AI Agents Compared

0:00Matt Anne, there are so many always on agents now. Dots Grockbot. My mis Gemini Park. I can't tell them apart. That's a fair complaint. None of these is a new announcement today. They're all out or rolling out. So, we're going to compare them, what each one is, where they overlap, and where they really

0:20differ. Wait, doesn't tech GPT already do stuff in the cloud for you? It does, and that's part of why people get confused. chat GPT work codeex cloud and cloud codes cloud sessions also run jobs in the cloud. So we'll compare those two plus cloud codes remote control. Okay, let's get into it. Here's the map I'll use. Group one is agents that run on a cloud computer and

0:51keep working between your conversations. Dots Grockbot, Mayus, and Gemini Spark. Group two is features where you hand over a job and it runs in a cloud workspace. Chat GPT work [music] codeex cloud and cloud codes cloud sessions. Remote control is a third thing because it doesn't move the work to the cloud at all. Isn't a scheduled job basically always

1:15on too? Good question [music] and it's why I'm not drawing the line at can it run while I sleep. Work can repeat on a schedule or trigger. Claude codes routines run as cloud sessions and Grockbot routines in Gemini Spark schedules do the same. The line is what you're talking to. A dot, [music] a bot, Mayus or Spark is one ongoing agent that remembers context and keeps a job going. Work codeex cloud and cloud sessions are jobs or sessions you start. That's my reading of how the docs

1:48describe them. First up, [music] dots from OpenAI. The help center puts it this way. Your DOT is an always on agent in chat GPT that can take on ongoing responsibility and keep making progress between conversations. Between conversations. So, it works even when I'm not chatting with it. Yes. Open AAI's developer docs say it tracks progress, [music] works out what needs to happen next, and can decide for itself when to pause and wake up. [music] So you don't need a fixed schedule for every followup. It can also research your connected information in the background, but that

2:30research is read [music] only. So where does it do all that work? On my laptop. Now each dot has its own cloud computer and its own [music] browser. and your own computer stays separate unless you choose to connect it. [music] That connection starts turned off and you can connect only one personal computer at a time. Open AAI's docs [music] lists three connections and each does a different job. A messaging channel to talk to your dot, [music] a connected app or plug-in to use that service within its permissions, and your local computer to work with its files

3:07and apps while it's [music] available. Connecting a messaging channel like Slack doesn't also connect your inbox or your computer. What if the dot needs to log into a website? It sends you a sign-in request [music] and you enter your credentials in a private form that goes to the browser outside the conversation or you can hit take over and sign in yourself. Open AAI also says some especially sensitive tasks like changing a password are always done by you. And how does it

3:38decide what it may do on its own? An automatic review checks actions that could affect your accounts [music] or share information against your instructions. Your custom rules and built-in safety requirements. On top of that, a custom rule can say take action without asking. Take action if preapproved. Ask before taking action or hand off to you. Open AAI itself adds that a dot can make mistakes. So it says to check anything with an important impact. On privacy, open AI says content in business, enterprise, and edu workspaces isn't used to improve its models [music] by default. And on personal plans, you choose whether your

4:20dots, conversations, and work are. Disconnecting an app doesn't delete what your dot already got from it. For that, you have to reset the dot. Okay, so who can actually get one? OpenAI says your first DOT is included in Pro or Business [music] Premium at no extra cost. Here are the pro plans. Pro is $100, $200 or $500 a month. And on OpenAI's business pricing page, a business premium seat [music] is $100 a month build annually or $125 build

4:54monthly. Is there any catch? Conversations with the dot don't count toward your chat GPT usage limits, but tasks it starts in codeex or work count as usual. At rollout, pro users in the European economic area, Switzerland and the UK are excluded. Enterprise users can try the beta only after a workspace admin turns it on and it's off by default. For companies, OpenAI is also starting a limited

5:23enterprise pilot of specialized dots. each with its own ID and a specific [music] job in the organization and it's working with Microsoft so they can be managed through agent [music] 365 and where do I talk to it you create it in the chat GPT desktop app or on desktop web since mobile can't create one after that you can message or call it from chat GPT and [music] message it from Slack or teams texting is coming and for now it's a limited beta for some pro users in the It can't call you at launch and roll out is gradual, so access can take several

6:02days to reach an account. Next, Grockbot from XI. It's built around named bots, each with a job, its own conversation, and working context that develops over time. XAI's docs call a bot a durable AI teammate. That sounds like a dot with a different name. Close, but [music] two things differ. You can have many bots at once and they can work in parallel. And all your bots share one cloud computer with the same browser sessions, files, and command

6:35line credentials. Sharing one computer sounds risky. XAI says as much. It tells you not to put a credential or file on it if another bot on your account shouldn't use it. And it says each bot screen is a separate work surface, not a separate security boundary. The upside is handoffs. One bot can continue from work, another bot saved, and browser signins persist. So you usually don't

7:02sign in for each task. And it keeps running if I close my laptop. XAI's doc say Grockbot works from a persistent cloud computer. It can use a browser, command line, files, and connected tools without depending on your laptop remaining open. What about passwords and payments? The bot asks you to take over for a password or pass key, two-factor authentication, a capture, or a payment or identity check. You do just that step in the computer view and tell it to continue. Running commands on your own desktop is separate. [music] It needs

7:40your approval per command by default. There are also auto review rules that decide which actions stop for your approval. And a team admin can lock rules for everyone on the team. How do I make it repeat things? With skills and routines. A skill is a reusable set of instructions, [music] and a routine tells one bot when to run a workflow on a schedule or after an event, such as a Slack message or

8:07[music] a GitHub notification. One bot can own up to 50 routines. And the price, Grockbot comes through Cursor. Access is included with every paid individual cursor plan and the cursor team's plan and the computers [music] bots work on run in cursors cloud. You can also link an individual super grock, super grock plus or super gro heavy subscription. Usage resets weekly. The desktop app runs on Mac OS, Windows and Linux and there are [music] iPhone, iPad and

8:40Android apps. Mayus is Meta's. Meta calls it a personal AI agent that doesn't just answer questions, but does the work and says it was built so anyone can use it out of the box with no technical experience. What does does the work look like? You share a goal and Mayuse helps develop [music] a plan, then advances the work on its own. It can open a browser, fill out forms, and negotiate for you. For longer jobs, it keeps working after you close the app and comes back when something changes or it needs approval, like before sending an

9:20email or making a purchase. Where does it run? On a dedicated virtual machine, Meta calls Mayu secure VM with its own browser. Your data and the credentials for services you connect are stored there. A separate agent called Sentinel runs on the same machine. And Meta says nothing Mayos does reaches the internet unless Sentinel approves it.

9:44So my can't see my passwords. Meta says Mayos has no visibility into passwords or payment [music] methods. Credentials you share go into secure storage so my can use them without seeing them. It checks with you before sensitive actions. Shows an audit trail of what it has done and plans to do. And you can opt out of your interactions being [music] used to train Meta's models. For payments, Mayus can check out with Link by Stripe. [music] And Meta says it's the first AI agent

10:15covered by Link's purchase protections. Shop Pay and One Password support are listed as coming soon. Later this year, Meta also plans [music] Mayu's confidential VM, which would encrypt the whole VM with a key only you hold. So not even Meta can access it. How do I talk to it and what does it cost? In the Mayus app or directly in WhatsApp. It's rolling out in the US on iOS, Android, and Muse.ai with AI glasses coming. Meta says it's free for most of what people need with subscription plans for people who want

10:52to do more. Okay, so what do the paid plans cost? Here's the plan list from Meta's help center. Mayio is free with a usage limit. And if you hit it, you can upgrade or wait until the limit refreshes. The power plan is $20 [music] a month for 500 million Mayus tokens per week. The maximum plan is $100 a month for 3 billion tokens per week. The same article says Mayus and its subscriptions are in limited testing [music] and

11:23aren't available in all locations yet. It doesn't say how big the free [music] limit is, so how far the free tier goes isn't something Meta spells out. Gemini Spark is Google's. It started rolling out on May 19th, 2026, [music] first in the United States, and Google's page describes it this way. Give it a [music] task and it works in the background 24/7, even if your phone and laptop are turned

11:52off. Is it the same idea as the others? Same idea, different home. Spark is built around your Google account. Gmail, Calendar, Drive, Docs, Sheets, Slides, YouTube, and Maps connect natively. And those connections are off by default until you turn them on. It's organized as tasks, schedules, and skills. A task is the goal. A schedule is when it runs, [music] either at a set time or after an event, like a delayed flight. and a

12:22skill is how to do it. Does it use my own browser? It can do both. Spark can browse with a remote browser in the cloud or work through your desktop Chrome with permission. With your own Chrome, you need to keep the device and Chrome awake. And if you close the device, it might switch to the remote browser. Google calls Spark an experimental feature in early development and says your supervision matters. It's designed to ask for confirmation before sending communications, [music] modifying your data, making purchases, or submitting web forms, and it may ask you to take control for passwords or payment

13:01details. Who can use it? You need to be 18 or over, use a personal Google account, and have a Google AI Pro or Ultra subscription. Google's help page says it's not available with work or school accounts for now and not in the European economic area, Nigeria, Switzerland, or the UK. So, no using it with my work account. Google's own pages don't quite match on that. The help page says Spark isn't available with a work or school account for now, but the Spark page says it's available to pro and ultra subscribers over 18 in select countries, as well as select business users. [music] Neither page explains how those fit

13:45together. So, if you want it for work, [music] check which kind of account you'd sign in with. On prices, here's [music] what Google announced at IO on May 19th. AI Pro at $19.99 a [music] month and AI Ultra starting at $99.99 with a $100 tier and a $200 tier. In that announcement, [music] Spark was listed under AI Ultra only. Since then, it has widened. Google's current plans page lists [music] access to Gemini Spark under AI Pro 2, which matches the help pages pro or ultra

14:23requirement. any limits. Spark's usage [music] counts toward Gemini's compute-based limits, and you can have up to 15 tasks running [music] at once. A schedule won't run if 15 tasks are already running. Now, group two, starting with chat GPT work. Open AAI's help center calls it an agent designed for longer multi-step work and finished deliverables. Think documents, spreadsheets, presentations,

14:53reports or sites. Does it have a computer of its own like a dot? On web and mobile, work runs in the cloud. In the desktop app, if your plan and workspace allow it, [music] it can also use local files and desktop apps with your permission. Admins can control workcloud, work local, and codeex local

15:13separately. So, it's only on when I start it. you start it, but it can also repeat on a schedule or trigger or monitor for changes. Event triggered tasks can start from a new Gmail message, a new Slack channel message or GitHub pull request activity. Those event triggers are for eligible plus pro [music] business enterprise and edu users not free or go

15:39cloud. Codeex cloud is for coding. You choose GitHub repositories. Codec sets up a cloud environment and each task gets its own workspace. Tasks run on open AI managed computers so they can continue while your computer is asleep. Then you review the changed files and commit or open a pull request. Network access is set per environment. You can limit allowed domains and use network secrets so codeex can [music] reach HTTPS services that need

16:12credentials. What does it cost? work follows the same usage structure as codeex. [music] So both draw on the work and codeex allowance included in your plan. Plus and business standard seats [music] get only limited use of GPT6 Astra with optional credits after that. OpenAI notes that its pricing examples are based on coding [music] tasks. So

16:36work usage will vary by task. Which plans can run Codeex Cloud? OpenAI's codeex pricing page lists [music] codeex on the web starting with plus at $20 a month and pro includes everything in plus. Its feature table shows codeex cloud isn't available if you sign in with only an API key in workspaces. Codeex cloud access is managed separately from workcloud and

17:03codeex local. Anthropic docs describe a cloud [music] session as a clawed code session that runs on cloud infrastructure instead of your machine. In their words, the session keeps running after you close your laptop and you can check on it or steer it from any device. Where does it actually run? By default on anthropic managed infrastructure [music] with each session in its own isolated virtual machine or in your organization's self-hosted

17:35environment. Network [music] access is limited by default and your GitHub credentials stay outside the sandbox. How do I start one? From cloud.ai/code, the code tab in the cloud mobile app, the desktop app, or with Claude cloud in a terminal. You can pull a cloud session back into your terminal with teleport from the CLI. It's one way though [music] you can't push an existing terminal session to the cloud. Cloud sessions can also watch a pull request [music] and respond to CI failures and review comments. That's called autofix and it needs the clawed GitHub app

18:12installed on the repository. On plans, it's pro [music] max and team plus enterprise users with premium seats or chat plus cloud code seats. Sessions share rate limits with your other cloud usage. And Anthropic says there's no separate compute charge for the VM. Cloning and pull requests require GitHub. Then what's remote control? A different [music] thing. Anthropic docs say the web and mobile interfaces are a window into that local session. So your computer has to stay on and the cloud process has to keep running. Your code execution and file system access stay on your machine. So think of it as remote steering, not remote work. It's available on Pro, [music] Max, Team, and

19:00Enterprise, API keys aren't supported. And on team, and enterprise, an owner has to enable it first. Are those the [music] only two ways to work with Claudy code when I'm away from my desk? No. Here's the table from Anthropic Docs. It lines up six ways by what triggers the work and where Claude runs. Dispatch is a task you message from the Claude mobile app, [music] and it runs

19:24on your machine through desktop. Remote control drives a session that's already running. Channels push in events from a chat app like Telegram or Discord and run on your machine. The other three, Slack, where you mention it clawed in a team channel and it runs in Anthropics cloud. Self-hosted environments [music] which run on your organization's own infrastructure on team and enterprise plans. and scheduled tasks which can run in the CLI in

19:54desktop or in the cloud. Controlling my own machine from a phone sounds like a security question. There's a setting for it. Trusted devices is in beta and off by default. On team and enterprise and owner turns it on and on pro and max you turn it on yourself. Then each browser, phone or desktop app you use for remote control has to be enrolled and your sign-in has to be no more than 18 hours old. You refresh it with Face ID, Touch ID, Windows Hello or a pass key. Anthropic also says it never receives or stores fingerprints, [music] face data, or any other biometric information. Only the

20:35devices public key and basic metadata are stored. So, which one for what? If your machine will stay on and you want your local files, tools, and MCP servers, use remote control. If the work should carry on with your machine off, use a cloud session. Let's line the four always on agents up. What they share? Each works from a cloud computer instead of your laptop and

21:03keeps going when your laptop is closed. [music] Dots, Grockbot, and Spark say so directly, and Mayu says it keeps working after you close the app. They share a safety habit, too. Sensitive steps come back to you. Dots leaves things like password changes to you. Grockbot hands over passwords, two factor captures, [music] and payments. May asks before emails and purchases, and Spark asks for confirmation before purchases and

21:30sending messages. Okay, so where do they actually differ? First, how many dots gives you one dot to start and Open AI says you'll be able to add more later. Grockbot [music] is built around a roster of bots working in parallel. Second, the computer. Grock bots share one computer. Each dot has its own cloud computer and Mayuse runs on a dedicated VM watched by a separate Sentinel agent. Third, where each one plugs in. Dots reaches more than 4,000 [music] apps through plugins and talks

22:05to you in chat GPT, Slack, and Teams. Spark is rooted in Google's apps. Mayuse works in its own app and in WhatsApp with link by Stripe for payments. Grockbot uses connectors plus a general browser and the plans. That's the fourth difference. Dots comes [music] with pro or business premium, Grockbot with cursor plans, or a super Grock link. Spark with Google AI Pro or Ultra and Mayos is free with a usage

22:35limit in the US. So why use Dots if I already have work? Work codeex cloud and cloud sessions start from a job you give them. A dot is the one Open AI describes as continuing between conversations, deciding when to wake up and follow through. This part is my read of the docs, not official advice. Start with where your work already lives. I mostly write code

23:02in GitHub repos. Then the taskbased [music] cloud options fit codeex cloud if you're in chat GPT or claude code cloud sessions if you're in cloud since both work from GitHub repositories [music] and end in a pull request. If you want to keep using your own machine from your phone, [music] that's remote control. I want something that tracks ongoing

23:24projects, not one-off tasks. That's what OpenAI's examples for dot show. Keeping an off-site plan on track, keeping a sales proposal current as the seat count changes, updating a study as new results arrive. Grockbot's own advice for a first handoff [music] is a real task across a few tools with a clear finish line, like a weekly

23:47pipeline review pulled from your CRM. Gmail, [music] then Spark. Google's examples include summarizing your inbox, turning a group trips email chains into a spreadsheet of receipts [music] and an email to the group, and setting recurring reminders from past invoices. And if I just want an agent for everyday life without paying, [music] is the one Meta says is free for most of what people need. Though the free tier has a usage limit [music] and the paid plans are $20 or $100 a month. It works in WhatsApp and it's US only so far. One caution. These agents can act in your accounts. So how much they ask before

24:34acting matters as much as what they can do. Dots and Spark both warn that they can make mistakes. So review anything important inside one company. Yes. A dot can create cloud tasks in a codeex cloud environment [music] you've already set up and on a connected computer it can create work or codeex tasks and cloud code's own bridge is the cloud and teleport pair we covered earlier endocross companies can dots go quad code for the always on agents no the docs for dots grockbot may and spark don't describe an official link to another company's agent

25:19There's one [music] and it's encoding. Openai publishes an open- source plugin on GitHub called Codeex [music] plug-in CC that lets you use codeex from inside cloud code. Here's its command list. /codex review is a normal readonly codeex review. slashcodeex adversarial review is a steerable challenge review and slashcodeex rescue hands a task to codeex. There are also commands to transfer a clawed code session to [music] codeex

25:50and to check or cancel background jobs. To use it, you need a chat GPT subscription free tier included or an open AI API [music] key and usage counts toward your codeex limits. It runs [music] through the codeex install on your own machine. The README also warns about an optional review gate [music] which has codeex review cla's response before it finishes. It says that can create a long running loop between the

26:18two and drain usage limits quickly. Beyond that plug-in, [music] what exists is shared plumbing first? You can message a dot in Slack. A Grockbot routine can start from a Slack message and a work task can start from a new message in a Slack channel. So, what could be the meeting point? In principle, whether one agents own post counts as a trigger for another agent isn't [music] something these docs

26:45say, so I wouldn't promise it works. GitHub is on firmer ground. Codeex Cloud can open a pull request, [music] and Claude's autofix works on any existing PR. You paste the PR URL into a session and ask it to autofix. So a PR from codeex could be handed to Claude [music] as long as the Claude GitHub app is installed on the

27:09repository. It's not an official integration, but both halves are in the docs. One last [music] thing, each agent you connect is another place that can act in your accounts. Grock bots, [music] for example, share one set of browser login. So check each product's approval settings before you link them. That's my [music] advice, not an official one. So the short version the always on agents dots grockbot mayus and spark run on [music] cloud computers and keep

27:42working between conversations. Work codeex cloud and cloud codes cloud sessions are jobs you hand over and remote control steers a session on your own machine. The real differences are how many agents you get, [music] whether they share a computer, where they plug in, and which plan you already pay for, and across companies. The bridges that rest on the docs are OpenAI's codeex plug-in inside Cloud Code and passing a pull request through GitHub. The official pages we used are listed in the

28:15description. Here's a question for you. Which chore would you hand to an Always On agent first? [music] Tell us in the comments. If this helped, please hit like and subscribe, and we'll keep comparing the AI tools you're choosing between.