When another agent is measuring latency or CPU, take your own heavy processes off the machine first, or its before/after numbers are noise
Two actors shared one laptop: an engineer agent was timing 'name end to end latency and idle CPU, before and after a change' with a harness that opens the webcam, runs YOLO on MPS, whisper on CPU and plays speech through the speakers, while the human's demo app (same camera, same models, same mic) was also running. A shared webcam can only be opened by one process, and two model stacks halve each other's throughput, so the measurements would have been wrong in a direction nobody could predict (the 'after' could look worse or better depending on who was busy). Before waiting on such an agent: (1) list what is running (ps -axo pid,etime,command | grep -E 'app\.main|bridge|latency') and stop your own app, test loops, speech tests and bridges, with SIGTERM then SIGKILL on the specific pid, never pkill on a broad pattern that could hit a teammate's process; (2) leave the other agent's runner alone and tell the human what you stopped; (3) wait cheaply: a background loop that polls git fetch and exits when origin/main moves (for i in $(seq 1 90); do git fetch -q; [ "$(git rev-parse origin/main)" != "$BASE" ] && break; sleep 30; done) gets you a completion notification without spending tokens, and re-arm it after each commit; (4) set a deadline for waiting (here: cut off at T-1h45 and test what is on main) because the last queued feature is the first thing to sacrifice. macOS has no timeout command; use the loop's own counter. Evidence: macOS, 8 cores, camera + MPS + whisper contention seen as ~10 s name latency earlier the same night.