How to Schedule AI Agent Tasks (Recurring, One-Time, or Now)

This is the how-to for scheduling a specific AI agent to run on its own -- the mechanics, not the philosophy. Whether unattended agents are where small business is heading is a good question, and autonomous agents for small business argues it; this guide assumes you have decided yes and want to know which buttons to press.

On Deelo, any agent you have built in the AI Assistant can be put on a schedule: recurring, one-time in the future, or right now. It runs the task you give it and reports the results back in its own chat thread. That is the whole feature, and it turns a chat-when-you-need-it agent into a standing coworker that handles work on a rhythm -- the Monday-morning pipeline review, the nightly lead follow-up, the weekly stale-deal sweep. Here is how to set it up.

There are three timing options, and picking the right one is most of the setup. Recurring is the workhorse: the agent runs on a repeating cadence -- every morning, every Monday, the first of the month -- for standing work you want handled without asking. One-time schedules a single run at a future moment, useful for a task tied to a date, like a follow-up sweep the day after an event. Run-now kicks off a run immediately, which is how you test a scheduled task before you commit it to a cadence.

Most standing agents are recurring. The trick is matching the cadence to the work: an invoice-chaser fits a daily or weekly run, a pipeline review fits Monday mornings, a monthly report fits the first. Do not over-schedule -- a task that only produces something useful once a week does not need to run every hour.

Scheduling is a deployment choice on an agent that already exists, not a way to build one, so build first. The agent needs its role, instructions, app access, and permission grid set the way any agent does -- a schedule just decides when that agent runs, not what it is allowed to do. If you have not built one yet, how to build a custom AI agent is the step-by-step.

This order matters because a scheduled agent runs unattended, so its permissions carry more weight than an agent you are watching in real time. Get the grid right first -- especially the verbs that touch customers or money -- then schedule it. An agent you would trust running while you watch is the only kind you should trust running while you sleep.

In the agent's settings, set its schedule. Pick recurring and choose the interval -- daily at a set time, weekly on a chosen day, monthly on a date -- or pick one-time and set the moment, or run it now. Match the timing to when the work actually needs doing: a follow-up agent should run when leads are fresh, a report agent should run before you need the report, not after.

Think about time of day too. A scheduled run that prepares customer messages is most useful finishing just before your workday starts, so the batch is waiting when you sit down. A cleanup or reporting task can run overnight when nothing else is happening. The cadence is not just how often -- it is when, relative to when you will act on the results.

A scheduled agent needs a clear, bounded task for each run, and this is where scheduled agents succeed or drift. Write the instruction to be self-limiting: 'chase invoices overdue by more than seven days that haven't been contacted this week,' not 'chase invoices.' The first has a clear boundary and will not re-nag the same customer every day; the second is a recipe for a run that does too much or repeats itself.

Think about what should be true before and after each run. A well-written scheduled task knows what it already did last time -- it looks for the work that is genuinely new since the last run rather than redoing the same set. You express that in plain language in the task, and it is the difference between an agent that quietly handles a queue and one that annoys the same people on a loop.

After each run, the agent reports what it did in its own chat thread -- what it looked at, what it decided, what it prepared or completed, and anything it flagged for you. This is your window into an unattended worker, so use it: skim the report the way you would a end-of-shift note from a staff member.

Early on, read every report closely -- it is how you learn whether the agent's judgment matches yours and whether the task is scoped right. Once it has run clean for a while, you can skim for exceptions. The report-back is also where anything waiting on you shows up: drafts queued for approval, items the agent deferred because they hit a guardrail. It is not just a log; it is the handoff.

Unattended is exactly when guardrails matter most, and on Deelo they do not switch off just because no one is watching. The high-risk floor still holds: external sends to customers, money movement, bulk changes, and sensitive-data access stay approval-required even on a scheduled run. So a nightly review-request or invoice agent does its thinking and prep unattended -- decides who, drafts the messages, sorts the queue -- and the actual outbound sends wait in an approval queue for your morning glance.

That is the honest and genuinely useful shape of a scheduled agent: it does the overnight work so you wake up to a prepared, reviewed-in-one-click batch instead of a blank slate, not to a pile of messages it fired at customers while you slept. And on any public channel, an anonymous user can never approve a high-risk action, so a scheduled agent facing the public defers and reports rather than acting. The schedule changes when the agent works, never what it is allowed to do without you.