Five AI sessions outweigh 100 short requests

Google’s realtime-agent example shows five 20-minute sessions can impose more committed load than 100 requests that each finish in 50 milliseconds. Its session-aware load-balancing pattern combines application-reported active sessions with CPU or memory pressure, rather than trusting request rate or open connections alone. For teams applying queue-level fixes to streamed agents, the operational implication is to route and benchmark by session duration, idle-to-active ratios, disconnects, and p95 startup latency before a seemingly quiet backend becomes a hot shard.