As developers increasingly orchestrate multiple coding agents in parallel, the ancient question of whether a process is 'running' has given way to a far more demanding one: is it actually working, waiting, or done? Traditional process monitoring, built for a simpler era, cannot distinguish between a productive agent and one silently awaiting human approval — a gap that, at scale, misdirects attention and fractures workflow. The emerging answer lies not in faster timestamps but in semantic time and multi-signal state machines that treat each turn of work as its own unit of truth, and completion