What Is a Proactive AI Agent, and How Do You Set One Up?

A reactive AI assistant is the one almost everyone has already used: you open a chat window, type a question or a task, and the assistant answers inside that same conversation. The moment you close the window, it stops existing until you come back and start again. Nothing happens while you're away, because nothing is scheduled to happen.

A proactive agent flips that. It starts a run on its own, at a time you chose in advance, whether or not you're watching. That single difference, who starts the session, is the real line between the two, not how smart the underlying model is. A short script that checks a folder every morning at 7 a.m. and sends you a one-line note is proactive. A state-of-the-art model sitting in a chat window, waiting for your next message, is not, no matter how capable it is in the moment you're using it.

Four traits tend to separate an agent that's genuinely proactive from one that's just been set to repeat the same action over and over. It runs on a schedule without a human clicking start each time. It remembers something from its last run instead of beginning from zero every time. It checks its own output through a separate pass before handing you the result, the way a person reads back over an email before sending it. And it asks before doing anything that can't be easily undone, rather than assuming it has permission because nobody told it to stop.

None of those four traits require a research lab. They're a design choice you make when you write the instructions for the schedule, whichever tool ends up running it. The rest of this guide walks through each one, then gets specific about which tools can run a schedule for you today.

You don't need to build anything from scratch to try this. Several tools you may already have access to can run on a schedule today, and it's worth naming each one plainly rather than treating "AI agent" as one single product.

Anthropic's Claude desktop app has a scheduled-tasks feature: you give it a task and a cadence, such as hourly, daily, weekly, on weekdays, or run-it-yourself, and it runs that task as its own separate session each time, able to use whatever connectors and files you've given it access to. As of late October 2026, Anthropic's own support documentation states that scheduled tasks run on their chosen cadence even while your computer is asleep or the desktop app itself is closed, with one exception: a task that specifically needs your local files or local apps only runs while your computer is awake and the app is open, because that part of the work has nowhere else to happen.

OpenAI's ChatGPT has a comparable built-in feature, also called tasks, reachable from a dedicated Scheduled section in the sidebar or by simply asking ChatGPT in conversation to check on something later. Per OpenAI's own help documentation, as of late October 2026, a task can run once or on a recurring cadence, and ChatGPT can also watch for a change and tell you only when something meaningful shifts, rather than reporting back on a fixed clock regardless of whether anything happened. Exactly which cadences and how many active tasks you get depends on your plan, so check the Scheduled section in your own account rather than assuming a number from this guide.

And before either of those existed, your own computer already shipped a scheduler. macOS and Linux both run cron, a decades-old system where you write one line stating the minute, hour, day, month, and weekday a command should run, and the operating system fires it exactly on that cadence, forever, with no AI company involved at all. macOS also offers launchd, which Apple's own developer documentation describes as the preferred way to schedule a timed job on the Mac, using a property list file with a key such as StartCalendarInterval to name the exact time. Windows ships Task Scheduler, with both a graphical Create Task dialog and a schtasks command you can run from the command line. None of these three require any AI product at all: they just reliably start a program at a time you choose, which is exactly the one trait every proactive agent in this guide actually needs.

Start with one narrow, low-stakes task, not an ambitious one. A good first task reads something and reports back; it doesn't send, spend, or delete anything. Summarizing an inbox, checking whether a shared document changed, or flagging a deadline are all better first tasks than anything that touches money or another person's inbox directly.

Write the instructions the same way you'd brief a new hire on their first day: what to look at, what to ignore, what counts as worth telling you about, and exactly what not to do on their own. The fourth part, what not to do, is the one people skip and the one that matters most once the task is actually unattended.

Pick a cadence that matches how often the underlying thing actually changes, not the shortest interval the tool offers. An inbox worth checking hourly is rare; a project board worth checking daily is common. A tighter schedule than the task needs just produces more reports you'll stop reading.

Let it run a handful of times while you're still watching closely, the way you'd check in on a new hire's first few days rather than disappearing immediately. Read every report for the first week. Only widen what the agent is allowed to do, or how long you go between checking on it, once you've seen it behave the same way on an ordinary day as it did on the day you set it up.

Picture a Monday morning agent built around the four traits this guide started with. It wakes up on its own at 7 a.m. because a schedule, not a person, started the session. It opens by reading a short note from its own last run, so it already knows what last week's board looked like instead of treating this Monday as the first time it's ever seen the project. It compares today's state to that note, drafts a short summary of only what changed, and then runs that summary past a second pass that checks it against the actual data before sending it to you, catching the kind of mistake a single unreviewed pass tends to miss.

That's memory, working toward a goal, and self-checking, three of the four traits, running quietly before you've had coffee. The fourth, asking before anything risky, mostly shows up by what the agent doesn't do: it doesn't email your team the summary on your behalf, it doesn't close out tickets it decided were resolved, and it doesn't touch the shared drive it has read access to. It hands you a report and waits, which is exactly the point.

The same shape works for triage and for reminders. An inbox-triage version of this agent reads everything that arrived overnight, sorts it by whether it actually needs you, and tells you about the three messages that do instead of the forty that don't. A reminder version checks, before it says anything, whether the thing it's tracking already resolved itself, and stays quiet when it has. In every version, the schedule is doing the same basic job: starting the session so you don't have to remember to.