On context management in long-running agents

Long-running agents face a fundamental tension: context windows are finite, but the work isn’t. The naive approach — keep stuffing the window until it overflows — fails predictably. The interesting question is what to do instead.

This is a placeholder; full content lands later.