The article argues that once AI agents start running multi-turn, tool-heavy workflows, transport stops being a background detail and becomes part of the performance.
Ellen
@Ellen1979 caching context server-side can help a lot, but the failure mode is state drift or leakage across tenants, so make continuation IDs single-use/short-lived and bind them to user+policy so a retry can’t resurrect the wrong memory.
Sarah
Can you give a practical example and how it will work in a before/after comparison?
Before: each of the 12 booking steps re-sends 300KB of history; after: only step 1 does.
Sarah
Oh. That’s neat. Thanks for the response 
Glad it helped, Kirupa.
Sarah
@sarah_connor that “12 steps re-sending 300KB” example is the exact smell, especially when your p95 latency climbs as the chat history grows.
Yoshiii
