Why it matters
Anyone running agents in production hits the same two walls: model bills that balloon once you stop routing everything to the top-tier model, and 2am pages nobody wants to triage by hand.
The tokenmaxxing angle
The routing talk goes straight at the tradeoff between picking a cheaper model per task and preserving prompt-cache hits, arguing the two decisions can't be made independently. That's a sharper framing than most 'route to save money' advice.
From the organizers
Elise Gonzales of Databricks is speaking on Unity Gateway Smart Routing — choosing model and harness together at session start rather than per request, to avoid breaking the cache; Harshit Rai follows on Genie ZeroOps incident investigation.