An agentic workflow can do a surprising amount on its own. It plans a task, implements it, reviews its own diff and runs the tests. The tempting conclusion is to remove people from the loop entirely. The better one is to put them exactly where they matter.
Where a person says yes
I keep a human at three gates:
- The plan. A wrong plan executed perfectly is still wrong. Reading a plan takes two minutes. Reverting a day of agent work takes longer.
- Anything irreversible. Merging to
main, deploying, sending something to a customer, deleting data. If it cannot be undone cheaply, a person confirms it. - The exceptions. When the evals disagree, when a check fails twice, when the agent is unsure. Escalation is a feature, not a failure.
Where they do not
Everything in between can run without approval: drafting, refactoring, writing tests, fixing lint. Asking a person to click "OK" a hundred times a day does not add safety. It trains them to stop reading.
Few gates, clearly placed, actually read.
That is the whole rule. The agent does the work. The human owns the decisions.