Automation is a design discipline
The interesting question in automation is not "what can we automate?" It is "where exactly should the human appear, and what do they need at that moment?"
We design workflows from that question. The technology is downstream.
Tools we reach for
- Temporal for long-running, durable, signal-driven workflows
- Airflow for scheduled data pipelines where the DAG matches the mental model
- n8n for integration-heavy automations that benefit from a visual editor and inspectable runs
- Hand-rolled state machines when no tool fits and the surface area is small
Designing the human checkpoint
A good automation has explicit hand-off points. The human sees what the system did so far, what it is proposing to do next, why it thinks that, and what happens if they say no. If those four things are not visible at the checkpoint, the checkpoint is decorative.
Failure modes we plan for
Idempotency for retried steps. Compensation for partially completed workflows. Replay for incident investigation. Dead-letter handling for events nobody owns yet.
Good automation makes the human's job feel calmer, not faster.