What Is an Always-On AI Agent?
What is an always-on AI agent? It holds a goal between sessions. Three architectures get sold under that label, and here is how to tell them apart.
An always-on AI agent is an agent that holds a goal across time instead of across a conversation. You give it an objective once, it keeps working on that objective when you are not looking, and it reports back. The label is doing a lot of hiding, though. Three quite different architectures get sold as always-on, they cost different amounts, they fail in different ways, and the one you want depends on whether your work arrives on a clock, on an event, or continuously.
The three things vendors call always-on
Kind | How it wakes up | What it costs when idle | Typical failure |
|---|---|---|---|
Scheduled | A timer, hourly or nightly | Close to nothing | Misses anything that happens between runs |
Event-triggered | A webhook, a new email, a file landing | Close to nothing, plus a listener | Event storms, or a missed event nobody notices |
Continuously resident | It never stops, it holds state and a session | Real money, continuously | Drifts, loses track, or silently stalls |
Most products marketed as always-on are the first two wearing the third one's clothes. That is not dishonest, it is usually the right engineering choice, but it changes what you can promise. A scheduled agent that runs every hour cannot answer a customer in ninety seconds, no matter how the landing page reads.
What actually persists between runs
The word always-on implies continuity, and continuity has to live somewhere. In practice there are four separable things an agent might keep, and a vendor may keep any subset:
The goal, which is the minimum. Without a stored objective there is no agent, only a scheduled prompt.
The history of what it already tried, which is what stops it repeating yesterday's failed approach.
Credentials and sessions, so it can get back into the tools it needs without you re-authorising.
A machine, meaning a filesystem and browser that survive between wakeups.
That last one is the newest and the most consequential. When OpenAI announced Dots it specified that each agent gets its own cloud computer and browser, which is a different product from an agent that borrows your session each time it runs. We covered the launch in detail in our report on OpenAI Dots.
If you want the mechanics of the memory layer specifically, how AI agents remember between sessions goes through the storage patterns.
Why always-on changes the risk picture
A chat session is bounded by your attention. An always-on agent is not, and that removes the safety property most people were unknowingly relying on: you were there.
Three things get worse:
Small errors compound. A wrong assumption in a five-minute task is annoying. The same assumption in something that runs for three days produces three days of wrong work.
Nobody is watching the boundary. An agent that gradually widens its interpretation of the task has time to do so.
Cost becomes a background process. Spend that accrues while you sleep is spend nobody approved.
This is why the serious implementations put an approval step between finished work and delivered work, and why the interesting question about any always-on product is not what it can do but what it does without asking.
How to tell which kind you are buying
Four questions settle it, and they are answerable from documentation:
What wakes it up? A clock, an event, or nothing because it never sleeps.
What is the shortest interval between the world changing and the agent noticing? If that number is an hour, it is scheduled.
Does it keep a filesystem and a browser, or rebuild from scratch each run?
What is the usage allowance, and what happens at the boundary? Vendors are often quiet here, and a stall mid-task is the common answer.
Terminology in this area is genuinely muddled, so it is worth being precise about the category before comparing products. We separate the neighbouring terms in agent versus assistant versus copilot, the broader concept sits in what agentic AI means, and the model-level background is in how AI models work.
FAQ
Is an always-on AI agent the same as an AI assistant?
No. An assistant responds when you ask. An always-on agent holds an objective and acts between your requests, which is why it needs stored state, its own credentials and usually an approval step.
Does an always-on agent cost money while it is idle?
It depends which kind you have. Scheduled and event-triggered agents cost almost nothing between wakeups. A continuously resident agent that holds a session is consuming something the whole time, and that is the architecture that produces surprising bills.
What happens when an always-on agent hits its usage limit?
Usually it stops, often mid-task, and often without a loud signal. Since several vendors have not published their allowances, this is worth testing deliberately on something that does not matter before you depend on it.
Do always-on agents need their own browser?
Increasingly they are given one, because sharing your browser profile means sharing your live sessions. An isolated browser limits what a mistake or a malicious page can reach.
How did this land?
About the author
Senior Editor, AI & Product
Cecilia leads the Swarmz editorial desk. She has spent a decade turning complex AI and product topics into writing people actually finish, and she owns the blog's quality bar.
