Alongside the late-May model launch came features aimed at continuous work inside the coding tool: a more intensive execution mode and workflows that adjust during the task.
The direction matches what other tools in the category took in the period, and it shifts the usage pattern from conversation to delegation. Instead of asking and evaluating each answer, the person sets the objective and the system works until it gets there.
The gain is real on long tasks, the kind requiring reading several parts of a codebase, trying, measuring and correcting. That's exactly the work consuming developer time without demanding creativity at every step.
The cost appears in two places. Consumption, because continuous execution makes many calls in sequence, and supervision, because what happens while nobody watches has to be logged to be reviewable afterwards.
The practice accompanying this kind of feature is familiar: bound the agent's reach in the repository, require confirmation before irreversible actions, and keep a record of what changed and who authorised it.