Duration is not autonomy
A task that can run for hours still needs boundaries. The useful shift is that an agent can maintain a plan, operate across many files, and keep testing as it works. Without checkpoints, longer runs simply produce larger surprises.
Leave durable breadcrumbs
Plans, small commits, test output, and concise progress notes make a long run inspectable. They also allow another agent or a human to resume after interruption. Hidden reasoning is not a recovery strategy.
Review by phase
Do not wait until a feature-sized diff appears. Review architecture before data migration, behavior before polish, and verification before release. Each checkpoint is a chance to correct course while the cost is still small.