Scheduled agent
- Agents
- Industry term
A scheduled agent is an AI agent that runs on a timetable, such as every weekday at 7 a.m., instead of waiting for someone to type a request.
The scheduler starts the run and supplies a saved task definition. The agent then needs access to current sources, permission for its actions, and a destination for its results. Because no fresh request is required, missing or failed runs need a way to become visible.
Where you’ll meet it
Each weekday at 7 a.m., an agent searches the news for your industry and emails you the five stories it ranks highest. This morning three of the five cover the same product launch from different sites, and a new regulation that affects your business didn't make the list.
What to do
Tell the scheduled task which stories matter most, such as regulations and competitors' prices, and count one event once even when several sites cover it. Ask it to record which stories it considered and why it selected each item. Check a few digests against trusted sources before relying on the ranking.
Easy to mix up
A cron schedule only sets when a separate job starts. A scheduled agent also decides what happens in each run, working that out from what it finds.
Check yourself
Your Monday agent runs the same instructions each week to email a sales summary. This week's uses different headings and a different order from last week's, though the figures check out. What does that tell you?
Go deeperScheduled runs need duplicate protection and missing-output alerts
A retry or catch-up can start the same work twice. Give each intended run an identifier and make the sending or writing step reject duplicates with that identifier; a date alone is insufficient if several runs are intended that day. A job set for the hour when daylight saving time changes may run twice or not at all, depending on the scheduler.
An agent's own error report cannot cover a run that never starts. A separate check for an expected email or file by a set time can reveal that missing run as well as work that started but never produced a result.
A task whose mistakes are easy to spot and undo, such as a summary only you read, makes a good first job for a scheduled agent, before it gets permission to send or change things for other people.
Sources
- Stripe API: Idempotent requests, docs.stripe.com
- Cloud Scheduler: Cron job format and time zone, docs.cloud.google.com
- Google SRE: Monitoring Distributed Systems, sre.google