it’s really interesting to see how human-like these agent harness system are designed towards. for example, opus 4.7 would plan for a task with bounded time-window and implicit effort-level, which are human best practices on task management.