Scheduled AI Agents Explained: How Always-On Automation Actually Works (2026)

Grok Bot's August 11, 2026 launch put "always-on AI agents" in front of a much wider audience, and a lot of that coverage focused on the architecture — each agent getting its own cloud computer. That's a real and interesting story, and our sister site covers it in depth. This guide is about a different question: once you have an agent that can run without you present, what actually needs to be true for that to work well in practice?

What "scheduled" means for an agent

A chat assistant only does something when you type into it. You open the window, ask a question, get an answer, and the session effectively ends when you close the tab. A scheduled agent inverts that: it starts itself based on a trigger, does a defined piece of work, and delivers the result somewhere you'll see it — without a human present at the moment it runs.

There are three common trigger types:

Cron. Runs on a fixed clock — every morning at 7am, every Monday, every hour. Good for anything that should happen on a schedule regardless of external events: a daily competitor brief, a weekly report, a recurring data pull.
Email triggers. The agent wakes up when a message lands in a specific inbox and reacts to it — drafting a reply, extracting data, filing something. Good for reactive work tied to correspondence.
Webhooks. Another system calls the agent the moment something happens elsewhere — a form submission, a payment event, a status change in another tool. Good for event-driven work that shouldn't wait for the next clock tick.

"Always-on" is a related but distinct idea — an agent that stays running continuously rather than waking up on a trigger and going back to sleep. Grok Bot's pitch leans on that framing: agents with their own cloud computer that keep working. Both patterns solve the same underlying problem — work getting done without you having to be the one to start it — but they have different cost and predictability implications, which is the next section.

Why billing model matters more than people think

When you're the one clicking "send" on every request, a metered per-call price is easy to reason about — you feel each cost as you incur it. That intuition breaks down completely once an agent is running unattended.

An agent working while you sleep is still billing while you sleep. If it hits a retry loop, encounters a page that keeps timing out, or just does more work than expected on a given run, nobody is there to notice and stop it in real time. This is exactly the cautionary story that came out of Grok Bot's early beta: one tester publicly reported that a single month of always-on agent use consumed more tokens than they'd used in the previous five years of ordinary interactive LLM usage. That's not a criticism of the product's capability — it's a direct consequence of metered billing applied to unattended, continuously-running work. Grok Bot's subscription bundle (roughly $200/month) includes an allotment of tokens and tool calls, and once that's used up, every additional token and tool call is billed on top, with no published ceiling.

There are two fundamentally different billing shapes for agent platforms:

Metered. You pay for what's actually consumed — tokens, tool calls, compute time — on top of any base subscription. This is fair in the sense that light users pay less, but it means your maximum monthly spend isn't knowable in advance. A misconfigured loop or an unusually heavy workload can produce a bill that surprises you after the fact.
Hard-capped. The platform sets a monthly ceiling and the service pauses when you hit it, rather than continuing to bill. You know your maximum possible spend before the month starts. The tradeoff is that if you consistently hit the cap, you either upgrade or the work stops until the next billing cycle.

Neither shape is universally "better" — a metered model can be cheaper for genuinely light, occasional use, and a hard cap can feel restrictive if your workload is spiky. But for unattended agents specifically, where nobody is watching the meter in real time, a hard cap converts an open-ended financial risk into a fixed, budgetable cost. That's a meaningfully different risk profile, and it's worth weighing more heavily than the headline subscription price when you're choosing a platform for scheduled or always-on work.

How approval gates work for agents that act without you watching

The other operational question that matters once an agent runs unattended: what happens right before it does something irreversible? Sending an email, posting a public reply, submitting a form on a customer's behalf — these are the moments where a mistake is hardest to undo.

An approval gate holds that specific action for a human to review — typically a quick approve or reject — before it actually goes out, while still letting the agent do everything leading up to that point autonomously: research, drafting, data gathering, analysis. The agent does the work; a person confirms the moment it becomes irreversible. That's a meaningfully different design from an agent that's fully autonomous end-to-end, and it matters most for exactly the always-on, unattended pattern this article is about — because nobody is reading over the agent's shoulder in real time, the gate has to exist somewhere in the pipeline instead.

Where that gate lives matters too. A gate buried in a dashboard you check twice a week is barely a gate — the point is to make the check fast and reachable wherever you actually are, not to add a step you'll forget about.

Delivery channels: where the result actually shows up

An agent that runs perfectly and delivers its output somewhere you never check is functionally useless. Delivery channel is an underrated part of choosing an agent platform:

Dashboard. Results land in a web UI you have to remember to open. Reliable, but only as useful as your habit of checking it.
Email. Familiar and easy to file or forward, but easy to bury in an inbox alongside everything else.
WhatsApp (or similar messaging). Results — and approval requests — land where a lot of people already have their attention by default, which tends to produce faster response times for anything time-sensitive, like an approval that's blocking a send.

The right channel depends on the task. A weekly report can comfortably sit in a dashboard. A same-day approval request usually shouldn't.

A buyer's checklist for scheduled agent platforms

What triggers does it support? Cron only, or also email and webhook triggers for event-driven work?
What's the billing model, and what happens at the ceiling? Does it pause, or keep metering?
Is there an approval gate before irreversible actions, and can you actually reach it quickly — including from your phone?
Is there an audit trail? Can you see what the agent actually did on a given run, or only the end result?
Where do results land? Somewhere you'll actually see them in time to act, or somewhere that requires a habit you don't have yet?

Where CloudyBot fits

CloudyBot runs specialists on cron, email, and webhook triggers, backed by a persistent, per-user isolated cloud browser rather than a machine shared across agents. Every plan uses a hard monthly cap — the service pauses at the limit instead of metering overages — starting at $19/month for Growth, with a free tier to try it first and no waitlist. Outbound actions like sending a reply wait for one-tap approval, reachable from WhatsApp on paid plans, and activity is kept in tamper-evident records with revision history so you can see what a specialist actually did on a given run.

Honest limitation: CloudyBot doesn't run xAI's frontier models and doesn't have native X integration. If those are specifically what you need, this isn't the right fit regardless of the billing model — see our honest look at whether Grok Bot is worth it for that side of the decision.

For the architecture story behind agents with their own cloud computer — the part this guide deliberately left out — read the architecture deep-dive on CloudAxis. And if you're specifically small-business-sized and weighing Grok Bot's real cost against what it needs to replace, see is Grok Bot good for small business.

Frequently asked questions

What does "scheduled AI agent" actually mean?

It means an agent that runs on a trigger — a cron schedule, an incoming email, or a webhook — instead of only responding when you open a chat window and type something. The agent starts itself at the right time, does the work, and delivers the result, without you being present to kick it off.

Why does billing model matter more than price for unattended agents?

Because an agent running while you're not watching can't self-limit its own spend. A metered model bills for every token and tool call, so a loop, retry storm, or unexpectedly heavy run can produce a bill you don't see until it's already happened. A hard-cap model pauses the agent at a known ceiling instead, converting an open-ended risk into a fixed, predictable cost — which matters far more for unattended work than the headline subscription price.

What's an approval gate and why do I need one?

An approval gate holds an agent's outbound action — sending an email, posting a reply, submitting a form — for a human to review before it goes out, instead of letting the agent act autonomously on anything it decides. It matters because an unattended agent that can act without you watching needs a checkpoint before anything irreversible happens, especially for actions that touch customers or public communication.

Cron, email trigger, or webhook — which should I use?

Use cron for anything that runs on a fixed clock, like a daily morning brief. Use an email trigger when the agent should react to something landing in an inbox. Use a webhook when another system needs to kick off the agent the moment an event happens elsewhere, like a form submission or a status change. Many real workflows combine more than one.

How does CloudyBot handle scheduled agents differently from Grok Bot?

CloudyBot runs specialists on cron, email, and webhook triggers with hard monthly caps that pause the service at the limit, and delivers results including WhatsApp on paid plans. Grok Bot runs always-on agents bundled into SuperGrok Heavy or Cursor Ultra (about $200/month) with metered token and tool-call billing on top. Both are scheduled or continuous automation — the difference is mainly in billing model and how work gets delivered back to you.