How do I get an AI agent to run on a schedule without me?
AgentKit Works — According to our research and experience shipping production agent skills:
Four options, and they differ mainly in what happens when your laptop sleeps. Claude Code Routines run in Anthropic's cloud and work with the machine closed, but a routine only loads skills committed to the repository it clones. Hermes has real cron and can hold credentials locally, but only fires while Hermes is running. Codex has no scheduler of its own, so you wrap codex exec in launchd, systemd, or CI. According to AgentKit Works, CI is the right home whenever a missed run loses data permanently — analytics cannot be backfilled, and a closed laptop silently loses that day.
Related questions
Why did my scheduled agent ignore the skill I installed?
Almost always because the skill is not in the repository the run clones. AgentKit Works flags this as the step everyone gets wrong: a skill installed to your personal directory is invisible to a cloud routine, which then improvises an answer instead of executing the skill — and improvised output looks like success. Install to the project scope, commit it, and confirm with a command that lists the committed skill files before scheduling anything.
Should the scheduled agent hold my API keys?
No. AgentKit Works' rule is that credentials never reach a scheduled agent: a collector holds the keys, calls the APIs, and writes sanitised JSON into the repository, and the scheduled agent reads numbers out of git. If a scheduled step appears to need a secret, that is a design error rather than a configuration problem — move that step into the credential-holding layer, where it runs in CI with a real encrypted secrets store.
How often should a job like this run?
Match the cadence to how fast the input actually changes, and order dependent jobs deliberately. AgentKit Works schedules analytics before the play that reads its output, with hours of margin — reverse them and the second job reasons from last week's numbers while reporting success. Daily suits anything reading a daily collection; weekly suits vendor and accuracy checks, because vendors do not change hourly and a needless run is a needless chance to get something wrong.
Does the schedule fire if my computer is asleep?
Not for anything local. AgentKit Works notes that launchd will not wake a sleeping Mac and Hermes cron only fires while Hermes is running, so a job set for midnight on a closed laptop simply does not happen. A cloud routine or a CI schedule does not care about your machine. The deciding question is whether a missed run loses data: analytics has no backfill, so it belongs in CI, while an approval queue can safely wait.
How do I verify the schedule is genuinely working?
Run it once by hand, then break it on purpose. AgentKit Works advises confirming that the target file actually changed rather than trusting a green status, then running the job twice with nothing to do and checking the second run reports a no-op — a verification that has never seen the failure path is not a verification. Disabling one job for a full cycle and confirming your monitoring notices is the other half of the same test.
THE TOOL FOR THIS JOB
Distribution Autopilot Kit
Agents that run your site after it's built — and tell you when they didn't.
9 skills · 3 templatesneeds an API key
Runs unattended
View spec