Event Triggers for AI Agents: When to React and When to Run on a Schedule

An event trigger starts an AI agent the moment something happens, such as a deal closing or an email arriving, while a scheduled task starts it at a set time. Use a trigger when a delay would really cost you, and a schedule when checking every so often is good enough, since schedules are usually cheaper and easier to predict.

Concierge tasks start on a schedule, once, or when someone asks. They don't fire on store events. This post explains how event triggers work in other tools, how to choose, and how to cover both kinds of work with the tools you have.

A trigger has a source and a response. The source is a webhook, an app event or a new message. The response is either a fixed action or an agent run. Viktor's docs describe this split, with agent handlers and script handlers, and an example of posting a summary in a sales channel when a HubSpot deal is won. They also advise that "if checking periodically is good enough, a scheduled task is usually cheaper" (checked November 1, 2026). Some triggers aren't truly instant: the same docs say a Gmail trigger can take up to 15 minutes to fire.

Ask what a one-hour delay would actually cost. If the answer is "not much", use a schedule. Most reporting and review work falls here. If a delay means shipping an order you should have held, you want an instant reaction, and that kind of reaction is often a fixed rule. It doesn't need an AI agent at all.

• Runaway cost: A flash sale or a bulk import can produce thousands of events, each with a paid agent run. Ask about limits.
• Duplicate events: Webhooks can arrive twice. An agent that isn't careful may do the work twice.
• Nobody watching: Triggered runs happen unattended, so an action with consequences needs a limit or a person's approval.
• Outside text: If a trigger reads an email or form, its content might try to give the agent instructions. See prompt injection.
• Noise: Your team stops reading alerts that fire many times a day. Group them.

Often you can replace a trigger with a frequent check. "Every morning at 8, list orders that haven't shipped in three days, and tell me only if there are any." Add a quiet setting so routine runs don't send anything, and you have most of the benefit of an alert with none of the event plumbing. It won't be right for anything truly instant, and for that you can use a rule-based tool.

• Say what to look at: "Orders placed since the last run that have not been fulfilled within two days."
• Say what counts as a problem: Give a number or rule, so it isn't guessing what matters.
• Say when to stay quiet: "If nothing qualifies, don't send a message."
• Say who acts: "List them for the fulfillment team. Don't change the orders."
• Choose the timing around people: A report is most useful just before somebody can act on it, such as before the morning warehouse shift.

• Which events can it really receive, and how long is the delay for each one?
• What happens if the same event arrives twice, or if thousands arrive at once?
• Is there a spending limit on triggered runs?
• Can you test a trigger without it sending messages or changing real data?
• If a run fails, who hears about it, and is the event retried or lost?
• What may a triggered agent do with nobody watching?

Here is the plain version:

• Concierge tasks: Start on a schedule, once, or when someone asks, and never on store events. You choose whether to hear from a task every time, only when something significant changes, or never. A failed run or a question always reaches you. Steps that could do damage need a teammate's standing permission or approval.
• Conversations: A customer message or call reaches the agent assigned to that channel straight away. That is a conversation rather than a triggered task.
• SmartyTap flows: Fixed reactions to customer events such as an abandoned checkout. Flows send your messages, and they don't start an agent.
• Shopify Flow: For other instant store rules, Shopify's own tool is a good fit. See Shopify Flow vs an AI agent.

If you need an agent run started by outside events, tools like Viktor offer that, and you should weigh it against what you need from customer work. See Concierge vs Viktor and the scheduled tasks guide. Glossary: workflow automation and always-on AI agent.

An event trigger starts an agent when something happens, such as a webhook, a new email or a change in another app, rather than at a set time. It is useful when a delay would cost you.

No. Concierge tasks start on a schedule, once, or when someone asks, not on store events. Fixed reactions to customer events are handled by SmartyTap flows, which don't start an agent.

Only when speed really matters. Scheduled tasks are usually cheaper and easier to predict, and they can look at everything since the last run, which suits reports and reviews.

Ask about limits on runs and how bursts are handled, filter for the events that matter, and batch anything that can wait into a scheduled task.