An always-on AI agent is an agent with four properties an ordinary chatbot lacks: a persistent identity you address by name, memory that survives between tasks, compute of its own that keeps running when your laptop is closed, and a way to start work without you (a schedule, a trigger, or another agent). Those four ship in three shapes. A managed cloud computer rents you the machine and runs the agent on it: Grok Bot, launched 11 August 2026, and Manus, whose Cloud Computer add-on was announced on 30 April 2026. A self-hosted framework hands you the code and lets you place it: OpenClaw and Hermes Agent, both MIT licensed. An own-hardware workbench runs agents on a machine you already have and gives you remote control of it: Continuum. "AI teammate" and "digital labor" are marketing terms for the same four mechanics, and neither implies autonomy, judgement, or accountability. The shape you pick decides who holds your credentials, whether idle time is billable, and whether you can leave.
- Four mechanics define the category: a persistent identity, memory that survives, compute of its own, and a trigger that is not you.
- Three delivery shapes, not a spectrum. Managed cloud computer, self-hosted framework, or own-hardware workbench. Each is a different answer to "whose machine".
- "AI teammate" and "digital labor" are positioning, not capability. Salesforce put digital labor into the vocabulary; xAI sells "a new kind of colleague". The mechanics under both are the four above.
- Persistence is the feature and the risk. Grok Bot documents one shared computer per account where "a login or file placed on the computer" is available to every Bot you own.
- Idle time is the hidden meter. A rented always-on machine bills whether or not it is working; Manus deletes the machine’s contents if you stop paying.
- The objection to own hardware is real and answerable. Remote control of your own machine is the third option, and it is the one nobody markets.
What "always-on" actually means
Strip the category down and there are four mechanics. A product is an always-on agent when it has all four. Miss one and you have something else with a similar name, which is why the comparisons in circulation keep talking past each other.
- A persistent identity. You address a named thing, not a session. "Ask Rosa to reconcile the invoices" is a different interaction from opening a new chat, because the name is a handle on accumulated state.
- Memory that survives the task. Not a longer context window: storage that outlives the turn. Files it wrote last week, a browser still logged into your CRM, a note about how you like the weekly report formatted.
- Compute of its own. Somewhere the work happens that is not the window you are typing into. This is the mechanic that lets the phrase "even when your laptop is closed" be true, and it is the one the whole category argues about.
- A trigger that is not you. A schedule, an inbound event, or another agent handing over work. Without this the agent is always-available rather than always-on, and the distinction is the entire product.
Notice what is absent from that list. Nothing about intelligence, nothing about autonomy, nothing about how good the model is. An always-on agent built on a mediocre model is still an always-on agent, and a brilliant model in a chat window is not one. The category is an architecture, not a capability level, which is why a page arguing about benchmark scores is answering a different question from the one you are asking.
It is worth being precise about the fourth mechanic too, because vendors are loose with it. A scheduled prompt that wakes a chatbot and emails you the answer satisfies the letter of it. xAI ships exactly that as a separate product, Grok Automations, alongside Grok Bot. What Bots add is that the scheduled thing operates real software with credentials attached, which is a much larger promise and a much larger blast radius.
The three delivery shapes
Every product in this category answers "whose machine does the work happen on" in one of three ways. The answer is not a matter of degree; it is a fork, and each branch has its own economics, its own failure mode, and its own kind of buyer.
| Shape | Whose machine | Examples | What you pay for | What you give up |
|---|---|---|---|---|
| Managed cloud computer | The vendor’s | Grok Bot, Manus Cloud Computer | A seat or credits, plus the machine, plus inference | The trust boundary, the model choice, and the exit |
| Self-hosted framework | Whichever you point it at | OpenClaw, Hermes Agent | A VPS or your own box, plus model API spend, plus your time | Setup, upkeep, and being your own operations team |
| Own-hardware workbench | A machine you already own | Continuum | Electricity, and inference on plans you already hold | A machine has to stay switched on |
Managed cloud computer
The vendor provisions a Linux machine, puts the agent on it, and sells you access. Grok Bot is the purest example: named Bots on a persistent cloud VM with a browser, a filesystem, and a terminal, reachable from a desktop or phone app. Manus reached the same place from the other direction, starting with a fresh sandbox per task and then announcing Cloud Computer on 30 April 2026, a persistent Ubuntu machine its own help centre describes as "a dedicated, persistent virtual machine in the cloud that allows Manus to run your applications, bots, and scripts 24/7".
The appeal is genuine and worth stating without sneering: there is no agent host or server to install or maintain, only the native client, nothing to keep awake, and nothing to debug at 2am. For most people buying an assistant to chase invoices, that is the whole value proposition and it is correctly priced.
Self-hosted framework
You get source code and you decide where it runs. OpenClaw is the incumbent here and appears in essentially every alternatives list written about this category. Hermes Agent, from Nous Research, is the other serious one: MIT licensed, with GitHub showing 234.2k stars on 22 August 2026, with a CLI, a desktop app across macOS, Windows, and Linux, messaging gateways for Telegram, Discord, Slack, WhatsApp, Signal, and email, and seven terminal backends (local, Docker, SSH, Singularity, Modal, Daytona, and Vercel Sandbox).
The appeal is control: your model provider, your data, your kill switch, no permanent account links. The cost is that you are now the operations team, and the churn is real. One commenter on the Grok Bot launch thread, thenbrent, put the failure mode plainly: "OpenClaw broke every update for a month, so I gave up and moved to Hermes."
Own-hardware workbench
The third shape is the least marketed and the most obvious once you see it: run the agents on a machine you already have, and reach it from anywhere. No VPS to rent, no vendor holding your logins, and no monthly bill for compute you already bought. Continuum is built on this shape, running Claude Code, Codex, Cursor, Grok, Gemini, and OpenCode on your own Mac or Linux box, with iPhone and web control of a session that is executing on your hardware.
This shape has one real requirement, and pretending otherwise is how vendors lose credibility: a machine has to stay switched on. If your only computer is a laptop that travels in a bag, the managed cloud shape genuinely wins and you should buy it.
"AI teammates" and "digital labor": the words and the mechanics
Two phrases dominate the category and neither one describes a capability. It is worth separating what they are doing rhetorically from what the product does mechanically, because the gap between the two is where buyers get surprised.
| Term | Where it comes from | What it actually denotes |
|---|---|---|
| AI teammate | xAI’s Grok Bot launch: "AI teammates you can give real work to", with the product page calling a Bot "a new kind of colleague" | A named agent with persistent memory that you address in a chat thread |
| Digital coworker | Press coverage of the same launch | The same thing, phrased for an executive audience |
| Digital labor | Salesforce, which built a whole product category page around it for Agentforce | Agent runs billed as a substitute for headcount rather than as software |
| Digital employee | Manus: "Let your digital employee take the night shift" | A persistent VM running scripts and bots on a schedule |
| Agent swarm | Practitioners, usually sceptically | Several agents in one thread passing work between them |
The teammate framing is doing real work, and not all of it is cynical. Naming a thing and messaging it is a genuinely better interface than a task queue, because it gives you somewhere to put context and a natural place to check in. On the launch thread h14h, who had been assembling the same setup by hand out of OpenClaw and Telegram bots, wrote that "Grok Bot simplifies the setup for this about as far as I imagine is possible, and frankly it’s a pretty slick experience." That is a fair verdict on the interface.
What the framing obscures is accountability. A teammate has judgement and can be held responsible. An always-on agent has neither, and the logs record its actions as yours. SilverBirch asked the sharpest version of this on the same thread: "Grok bot snags your credentials and then pretends to be you whilst surfing the web. If it uses your account to post on X, is it adhering to X terms of service or is that violating policy by using a bot?" Nobody in the category has answered that, and no product ships an answer today.
Treat "digital labor" the same way. It is a pricing frame, not a technical one: it invites you to compare the bill against a salary rather than against a software licence, which makes $300 a month look cheap. That comparison is only honest if the failure modes are also employee-shaped, and they are not. An employee who is unsure asks. An agent that is unsure proceeds, and eesel’s review of Grok Bot notes there is no dry-run mode: "A test run performs real work, it can navigate websites, change files and call connected tools." If you rehearse an email send, you sent an email.
What the shape decides
Five consequences follow from the shape, and none of them appear on a feature comparison. These are the questions to ask a vendor, in this order.
| Question | Managed cloud | Self-hosted | Own hardware |
|---|---|---|---|
| Who holds your logins? | The vendor’s machine does, in a browser profile you cannot inspect | You do, on infrastructure you chose | You do, on the machine already holding them |
| Is idle time billable? | Yes. The machine runs whether or not it is working | Yes, at VPS rates, unless you host it on something you own | No. The machine is already paid for |
| Can you choose the model? | Usually not. Grok Bot ships no picker and documents no default | Yes. Hermes switches provider with one command | Yes, including across labs in the same day |
| What happens if you stop paying? | The machine goes away. Manus states files, tools and databases on it "will be deleted" | Nothing. The code is MIT licensed and still on your disk | Nothing. The app keeps working with your own keys |
| Who is liable for what it did? | You are. It acted as you, and there may be no audit view yet | You are, but you hold the logs | You are, and the logs are on your machine |
The second row is the one people discover late. An always-on machine is always-on in the billing sense too. Manus prices its Cloud Computer as a flat monthly add-on (Basic at $10 a month for 2 vCPU, 1GB of RAM and 35GB of storage; Standard at $30 for 4GB and 70GB; Advanced at $50 for 8GB and 120GB) on top of a membership that starts at $20 a month for 4,000 credits. Grok Bot takes the other approach: the machine comes with the seat, and the meter runs on tokens instead, against a weekly allowance xAI does not publish and past it at raw model rates with no spend cap.
Both models can produce a bill you did not plan for, but they fail differently. A flat VM charge is predictable and wasteful. An uncapped token meter is efficient and unbounded. If you cannot tolerate a surprise, the flat one is safer; if you cannot tolerate waste, neither is, and the third shape is where you end up.
An honest map: who should pick which
The category rewards matching the shape to your actual constraint, and the constraint is rarely the one the marketing addresses. Here is the map as plainly as we can put it, including the cases where our own product is the wrong answer.
| If this is true of you | Pick | Why |
|---|---|---|
| Your only computer is a laptop that travels | Managed cloud | Nothing else can be always-on. This is the honest case for renting |
| The work is office admin in SaaS apps with no API | Managed cloud | Computer use on a hosted browser is the mature version of this, and Grok Bot is good at it |
| The work is code in repositories you own | Own hardware | Your machine already has the repo, the toolchain, the secrets, and the git history |
| You need to choose the model per task | Self-hosted or own hardware | Managed products route automatically and mostly ship no picker |
| You have a compliance review to pass | Self-hosted or own hardware | No published Grok Bot SOC 2, ISO 27001, GDPR or HIPAA claim was found in its doc set, and the audit view is described as coming |
| You enjoy running infrastructure, or need to | Self-hosted | Hermes and OpenClaw give you every dial. Budget the upkeep honestly |
| You want zero setup and one bill | Managed cloud | This is what the seat price buys, and it is a real thing to buy |
| You already pay for Claude, Codex or Cursor subscriptions | Own hardware | You have bought the inference already. Renting a second computer to use it again is paying twice |
Two rows deserve expanding because they are where most people land.
If the work is office admin, the managed cloud shape is not a compromise, it is the product that was built for you. A Bot that logs into a supplier portal, downloads a statement, reconciles it against a spreadsheet, and messages you the discrepancy is doing something the other two shapes handle worse. The trade you are making is that the vendor’s machine now holds a session for that supplier portal until the site or session expires, you sign out, or access is revoked, and in Grok Bot’s case shares it with every other Bot on the account. That is a defensible trade for a low-stakes portal and a bad one for your bank.
If the work is code, the calculus inverts. The repository, the toolchain, the environment variables, the half-finished branch, and the credentials your build needs are already on your machine. A cloud agent has to be given all of that again, which is why every product in the background-agent category has an environment configuration file and a setup script. Running the agent where the code already is skips that entire class of problem, and it is the reason background agents and always-on teammates ended up as two separate product categories rather than one.
The objection to own hardware, and the answer
The strongest argument against the own-hardware shape is not about capability. It was stated on Hacker News by phoghed, and it is worth quoting in full because it is the objection every vendor in this category is really selling against:
Three of those four sentences are correct. Most people will not self-host, self-hosting is available and unpopular, and wanting one agent reachable from every device is the right instinct. The part that does not follow is "locally", because it quietly assumes that running on your hardware means sitting in front of your hardware.
It does not. A machine at home or in the office with remote access is a third option, and it satisfies the actual requirement in the quote: one agent, running somewhere else, reachable from every device. The difference from a managed cloud computer is only who owns the box. Continuum is built exactly this way, which is why the iPhone and web clients are not an afterthought: a session runs on your Mac and you drive it from a phone on the train, which is functionally the same experience as a hosted agent and materially different on cost, model choice, and who holds the credentials.
The economics are not close once a machine is involved for more than a few months. A Mac mini starts at $599 in the United States for the base M4 configuration with 16GB of memory, paid once. A Grok Bot seat is $40 to $300 a month before the uncapped token meter. A Manus Cloud Computer is $10 to $50 a month on top of a membership. The stated rented options pass the price of the bought machine in roughly 2 to 20 months before electricity, depending on tier, and then keep going.
What the rented machine genuinely buys, and this is the part own-hardware advocates skip: it is on when your house loses power, someone else patches it, and you never think about it. If that is worth $40 a month to you, buy it. That is a real product, honestly priced, and we are not going to pretend otherwise.
What nobody in the category has solved
Whichever shape you pick, four problems are unsolved across all of them as of August 2026. A vendor claiming otherwise is selling.
- Prompt injection. An agent with a browser and your credentials reads untrusted text all day. Nobody has a defence that works at the reliability level this category needs; a mitigation that fails one attempt in forty is not a mitigation when the agent runs unattended a thousand times a week.
- Audit and reconstruction. Grok Bot’s documentation describes an audit view as coming. Its 20-record retention limit was reported by Daily Dose of Data Science and remains unconfirmed by xAI. If you cannot reconstruct what an agent did last Tuesday, you cannot answer for it.
- The dry run. There is no rehearsal mode in any mainstream product in this category. Testing is doing.
- Revocation. Deleting an agent does not necessarily remove the sessions it opened. On a shared machine, credentials outlive the agent that typed them.
None of this argues against using always-on agents. It argues for scoping them: give them accounts that can be revoked independently, keep them out of systems where an unattended mistake is expensive, and prefer shapes where you hold the logs. That is unglamorous advice, and it is the advice that will still be right in a year.
If you want the same analysis applied to one product rather than the category, what Grok Bot actually is reads the documentation against the marketing, the security and compliance page works through the buyer objections, and the alternatives comparison ranks the field on merit.
Questions people ask
What are always-on AI agents?
An always-on AI agent has four properties a chatbot does not: a persistent identity you address by name, memory that survives between tasks, compute of its own that keeps running when your laptop is closed, and a trigger that is not you (a schedule, an event, or another agent). Products with all four include Grok Bot, Manus, OpenClaw, Hermes Agent, and desktop workbenches that run agents on your own hardware. Products with three of the four, like a scheduled chatbot prompt, are usually marketed as if they had all four.
What is an AI teammate?
It is a marketing term for a named always-on agent you message in a chat thread. xAI launched Grok Bot on 11 August 2026 with the line "AI teammates you can give real work to", and its product page calls a Bot "a new kind of colleague". The interface metaphor is genuinely useful, but the word teammate implies judgement and accountability that the product does not have: the agent acts using your credentials, and every system it touches logs the action as yours.
What does "digital labor" mean?
It is a pricing frame that invites you to compare an agent subscription against a salary rather than against a software licence. Salesforce built a category page around the term for Agentforce, and Manus uses the sibling phrase "digital employee" in its Cloud Computer launch copy. The mechanics underneath are ordinary: a persistent machine, a scheduler, and a model. Judge the bill against what the work costs you today, not against a headcount.
Do AI agents need their own computer?
They need somewhere to run that is not your chat window, which is a weaker requirement than it sounds. That can be a vendor VM (Grok Bot, Manus Cloud Computer), a VPS or container you rent (OpenClaw, Hermes), or a machine you already own with remote access (Continuum). The differences are cost shape, who holds your logins, and whether idle time is billable. Do AI agents need their own computer works through the whole question.
What is the difference between an always-on agent and a background agent?
A background agent is per-task: you hand it a job, it runs somewhere else on an ephemeral machine, and it hands back a result, usually a pull request. Cursor Cloud Agents, Codex cloud tasks, and GitHub Copilot cloud agent all work this way. An always-on agent is persistent: the same named entity keeps memory, files, and sessions between jobs and can start work without being asked. Both run remotely; only one accumulates state.
Are always-on AI agents safe to give my logins?
Scope them rather than trusting them. Prompt injection is unsolved, dry-run modes do not exist in mainstream products, and audit trails are thin (Grok Bot’s documentation describes an audit view as coming). Give an agent accounts you can revoke independently, keep it out of banking and payroll, and prefer a shape where you hold the logs. Grok Bot specifically documents one shared computer per account where any login is available to every Bot you own, which rules out using one Bot as a sandbox and another as the trusted one.
Which always-on agent is cheapest?
Running agents on hardware you already own, because the compute is a sunk cost and you pay only for inference on plans you already hold. Among rented options, a Manus Cloud Computer is $10 a month at the Basic tier on top of a $20 membership, and Grok Bot starts at a $40 per seat Cursor Teams plan before an uncapped token meter. A Mac mini at $599 once reaches break-even against the stated rented options in roughly 2 to 20 months before electricity.
Can I run an always-on agent on my own laptop?
You can, but the laptop has to stay awake and connected, which is the honest weakness of the own-hardware shape. The usual answer is a desktop, a spare Mac mini, or a home server, with remote control from a phone or browser so you are not tied to the room the machine is in. If you genuinely have no machine that stays on, a managed cloud product is the right purchase and you should make it without guilt.
Sources
Every figure above was read from these pages on August 2026. Vendors reprice without notice; if you find a stale number, tell us.
- Introducing Grok Bot (xAI) 11 August 2026 launch, beta status, the "AI teammates you can give real work to" positioning
- Grok Bot overview (docs.x.ai) one shared cloud computer per account, per-Bot screens, the security-boundary warning, memory, routines, connectors
- Introducing Cloud Computer (Manus) announcement post dated 30 April 2026, the always-on Ubuntu machine, "Let your digital employee take the night shift"
- Cloud Computer plans and billing (Manus help centre) Basic, Standard and Advanced specs and prices, and what is deleted if you stop paying
- Hermes Agent (Nous Research) MIT licence, persistent memory, self-created skills, five to seven sandbox backends, messaging gateways
- NousResearch/hermes-agent on GitHub MIT licence, star count, install command, provider list, terminal backends
- What is digital labor? (Salesforce) the vendor definition that put "digital labor" into the category vocabulary
- Grok Bot launch thread on Hacker News the launch-day thread: the OpenClaw churn report, the shared-computer questions, the accountability objections
- Hacker News: "I very much don’t want to run it locally" phoghed’s comment, the clearest statement of the objection to own-hardware agents
- Grok Bot review (eesel) the compliance gap list, the missing audit view, no dry run, uncapped overage
- Mac mini (Apple) the base M4 configuration, 16GB memory and 256GB storage, $599 in the United States
- Continuum pricing the free app, and Plus, Max 100, Max 200 and Ultra hosted inference with weekly allowances