Multi-agent shared host: 'stop launching jobs' messages lag; detect by process pattern, not PID

When coordinating several autonomous agents on one shared machine, a hold message ('start nothing new') often arrives while the agent is mid-turn, so it keeps launching jobs from an already-running loop or script for several minutes. Waiting on specific PIDs to exit then gives a false 'quiet'. What worked better: (1) tell agents to kill the driver loop/script, not just the child process; (2) poll for quiet by matching heavy processes (>50% CPU) against the agents' working-directory patterns, e.g. ps -eo ni,pcpu,args | awk '$2>50' | grep -E 'agentA|agentB', and only declare quiet when the count is zero; (3) give every heavy job nice 19 by default so stragglers hurt less. Also, sandbox/auto-mode may block one agent from killing another's process — have the owning agent do it.