An always-on agent is a program that holds a goal, starts itself on a clock or an event, and acts without being prompted each time, then tells you what it did. The defining property is persistence, not round-the-clock computing. A chatbot waits for you. A cron job has no judgement. An always-on agent has both a trigger and a decision to make.
Part of the Always-on agents series, post 1 of 33.
I run a two-person company, and most mornings the first thing I read is a log written by software that worked while I slept. One job published a blog post at 09:05. Another checked the whole site at 09:40 and stayed quiet, which is its way of saying nothing broke. And Leo, the product I make, spent the night watching what people were responding to and booking clips onto a calendar. Only one of those three is what I would call an always-on agent, and working out why is the fastest way I know to explain the term.
What is an always-on agent?
The definition I use has three parts. An always-on agent holds a goal that outlives any single run. It starts itself, on a clock or when something happens, instead of waiting for me to open a window and type. And it acts and reports: it does something in the world and then tells me, or decides there was nothing worth doing and tells me that instead.
Everything else is detail. It might run on a Mac mini under your desk, on a five-dollar server, or on a cloud computer OpenAI will rent you by the month. It might use a large model for every decision or a small one for most of them. Those choices matter for cost and safety, and I cover them later in this series, but they are not what makes the thing an agent.
One line I keep coming back to, from a 2026 survey of these systems: the defining property is persistence, not literal continuous execution. My agents are idle almost all the time. What never stops is the loop: check, decide, act, record, sleep. If the Mac is asleep at 09:05, the loop still has to hold at 09:29 when it wakes. I will come back to that morning.
You will see the same idea under other names. OpenAI calls its version dots. Microsoft introduced Scout as your always-on personal agent and has since renamed it Autopilot. Google sells Gemini Spark as a 24/7 personal agent. Writers who prefer neutral terms say background agent, ambient agent or persistent agent. I use always-on because it names the property I care about and not the vendor.
What it is not: three things that get mistaken for it
I find the negative definition more useful than the positive one, because the mistakes are where the money and the risk go.
- A chatbot is not an always-on agent. It has judgement but no persistence. When I close the tab, it forgets the goal. Every session starts from zero and nothing happens between sessions. Most of what people call an AI assistant today is this.
- A cron job is not an always-on agent. It has persistence but no judgement. My 09:05 publisher is a cron job wearing a nicer coat: it runs on a schedule, checks what is due, uploads it, and never once decides anything I did not decide for it in advance. I like it that way.
- Workflow automation is not an always-on agent either. Tools like Zapier, n8n and Huginn have triggers and persistence, and they have been running unattended for a decade. What they lack is a decision that was not written down as an if-statement before the job ran. The moment you add a model to pick between outcomes, you have crossed into agent territory, and you should know you have.
My rule of thumb: if I can write the decision as a rule, I write the rule and I do not call it an agent. Agents are for the cases where the rule would be longer than the judgement, or where the inputs change faster than I can rewrite the rule.
The three tests I apply
When someone tells me they have an always-on agent, I ask three questions. The answers sort almost everything.
1. Does it hold a goal across runs?
Not a prompt, a goal. Leo's goal is that a channel keeps posting video that its audience responds to. That goal is the same at 3am on a Sunday as it is on launch day, and the agent carries state between runs so that it knows what it has already posted and what is booked. If the thing you are describing starts every run with a blank memory, it is not holding a goal. It is being handed one.
2. Does it start itself?
A clock counts. An inbound email counts. A webhook from GitHub counts. A person opening an app and typing does not, however clever what follows. This is the test the launches this autumn are really about: dots, Muse and Spark all run on a computer that stays on for you, so the agent can wake without you.
3. Does it act, and does it tell you?
Acting means something in the world is different afterwards: a post is live, a draft is in your inbox, a pull request is open, a calendar hold exists. Telling you is the half people forget. My site watchdog only emails when something is wrong, and the absence of an email is a signal I have learned to trust because I built the job to fail loudly. An agent that acts silently is a liability. An agent that reports without acting is a dashboard.
By my own tests, my publishing job passes two of three and fails the first: it holds no goal, it executes a list I wrote. The watchdog passes all three in a narrow way: it holds a goal (the site stays healthy), starts itself, and acts by emailing me. But every decision in it is a rule. Leo is the only one of the three that makes a judgement I did not write down in advance: which trend, which clip, and whether to post at all. Its rule for the last one is the part I am proudest of. No match, no post. Never filler.
What the big launches mean by the word
I read every announcement this autumn with the three tests next to me, and the vendors are closer to agreeing than their marketing suggests.
| Who | What they call it | What stays on | My reading |
|---|---|---|---|
| OpenAI | dots, inside ChatGPT (29 Sep 2026) | A cloud computer per dot, with your connected apps | Passes all three tests. The launch demos were ongoing jobs, not chats. |
| Microsoft | Scout, now Autopilot (2 Jun, renamed 25 Sep) | An agent with its own identity in your Microsoft 365 tenant | The identity is the interesting part: it acts as itself, not as you. |
| Gemini Spark (19 May 2026) | Google Cloud virtual machines, so tasks survive a closed laptop | Starts itself and acts; how much goal it holds between tasks is still unclear to me. | |
| Meta | Muse (8 Sep 2026) | A dedicated virtual machine holding the agent and your data | Built for long-running tasks. The trust question is the whole story. |
The common thread is not the model. It is that every one of them ships a computer that stays on for you, with an identity and a set of permissions attached. That is exactly what I had to build by hand for my own jobs, and it is why the setup guides in this series spend so long on the boring parts: the service account, the sleep settings, the restart policy.
Why persistence is the point, not uptime
On the morning of 21 September my publishing job did not run at 09:05. The Mac was asleep. It woke at 09:29, and the scheduler, catching up, fired the job before Wi-Fi had connected. The first call to the cloud failed on a DNS lookup and the run gave up. That day's post went out at 18:55, by hand, after I noticed.
Nothing about that failure was about compute. The job was never meant to be running continuously. It was meant to be persistent: to hold the goal of publishing everything due by today, survive an interruption, and finish the job on the next chance. So I changed three things, and they have become my rules for anything I let run unattended.
- Idempotent. Before acting, check whether the action already happened. My publisher asks the storage bucket whether the post exists before it uploads. A double run does nothing.
- Catch-up. Act on everything due on or before today, not only on what is due today. A day the machine is asleep heals itself on the next run.
- Wait for the world. The job now waits up to ten minutes for the network to resolve before its first call, and retries once after five more. An agent that fires into a dead network is not always-on, it is always-off with a schedule.
If you take one thing from this post, take that list. Every always-on guide in this series, on every operating system, is an elaboration of those three rules plus a fourth about identity.
Where I draw the line
There are things I will not hand to an agent, and the line is not about intelligence. It is about reversibility and visibility. I let agents publish, because I can unpublish. I let them draft, because a draft is not sent. I do not let them spend money without a cap I set, and I do not let them log in as me. Every job I run unattended against our cloud uses a service identity with the narrowest permissions that still let it work, because the default login on my machine expires, and an agent that silently loses its credentials at 3am is the kind of failure you only find in the morning.
That is also why I am wary of one pattern in the consumer launches: an agent that holds my email, my calendar, my payments and my passwords on one machine. It may be the right trade for some people. It is not one I have made yet, and the post on permissions in this series explains why.
What this series covers
One post a day for the next month, all written by me and all from things I have run. First the ground: who is shipping always-on agents, what they cost, and how an agent differs from the automation you probably already have. Then the builds: the use cases from the launch demos, rebuilt with free software. Then the setup guides: cloud, Mac, Windows, Linux, and one for each major distribution, each tested on a fresh machine or container before it goes up. Then the open-source projects and the free models. And at the end, the honest account of how my own company runs on these things, failures included. The whole series lives on one page, and if you want to see the agent I run for social video, it is here.
Sources
- OpenAI, DevDay 2026 recap (accessed October 20, 2026)
- OpenAI, Introducing dots (accessed October 20, 2026)
- Microsoft 365 Blog, Introducing Microsoft Scout: your always-on personal agent (accessed October 20, 2026)
- Google, Gemini Spark, your 24/7 personal AI agent (accessed October 20, 2026)
- Axios, Meta debuts Muse, its long-planned personal AI agent (accessed October 20, 2026)
- heym.run, Always-on agents: listening is free, thinking is not (quoting a June 2026 survey) (accessed October 20, 2026)
Written by Lera Leonteva. Vendor claims are quoted from their own announcements, linked above, as read on the access date.