When should a coding agent own the execution step?

A practical question about moving from model reasoning to verified repository execution.

I am trying to draw a clean line between “ask a model for the approach” and “let a coding agent execute.” The current rule of thumb: if the task has a repo, tests, and a concrete done condition, execution should move into the environment. If the task is still ambiguous, keep it in discussion until the acceptance criteria are sharper. Where do people usually put that boundary?