Agent Routines
Agent routines are scheduled prompts or tasks that let an agent perform a known workflow without an interactive session.
Definition
An agent routine is a prompt sent to an agent at a configured time or interval. The video gives a daily YouTube-to-Substack workflow as an example: a scheduled task invokes a custom skill, drafts an artifact, and leaves it for human review.
Execution Levels
- Local: desktop routines are simple to create but run only while the computer is on.
- Cloud or dedicated host: a VPS or always-on agent can run routines while the operator’s computer is off.
- Shared workspace: skills and memory must be synchronized when the routine runs on a separate machine; the video names Syncthing as one possible file-sync tool.
Routines should follow tested claude-skills and read controlled agent-memory-systems rather than inventing a workflow on every run.
Implications: Routines are the scheduling layer between reliable skills and unattended automation. Start locally with human review; move to an always-on host only when the task, inputs, outputs, and failure handling are understood.
Open Questions
- What monitoring, retry, and approval controls are required before a routine can run unattended?
- Which parts of the workspace may be synchronized to a cloud or dedicated agent without weakening provenance or access controls?
Related
- agentic-os — broader operating-system framing that includes routines among its components.
- claude-code — one agent environment described as a routine runtime in the source.