Here's the answer, on the line Anthropic's own guidance draws: a workflow follows a predefined path, step one, then two, then three, while an agent directs its own process. The operator's rule: anything describable as fixed steps becomes a workflow, and agents are reserved for the work that genuinely requires judgment. Workflows are cheaper, faster, auditable, and boring, and boring is the goal for the 80 percent of operations that repeat.
Three surfaces carry the automation: schedules for recurring work on a clock, triggers that fire on events, checks that run before a tool acts or after work completes, and headless runs invoked by other systems.
Building a pipeline is five steps: decompose the job into steps with named inputs and outputs. Assign each to a rail, a specialist, or a human gate, and money, sends, and judgment calls get human gates, always. Specify every handoff format. Define failure per step. And dry-run the whole thing in log-only mode until the log reads clean, because you automate the process that already works manually, never the chaos you hope automation will organize.
Then the supervision machinery that doesn't sleep: hooks that turn your governance page into enforced policy, a pre-action gate on untouchable territory and a post-action check running your acceptance criteria. One failure inbox for everything, checked daily, because ten scattered error locations equal zero. And a kill switch written and tested per automation, kept where panic can find it.
Ask the blast-radius question at build time: if this misfires for a week, what's the worst it does?
The full rails system is Volume 4 of our Advanced Operator series.
The dream isn't automation that impresses. It's automation so boring it disappears.