Optimize multi-agent performance and cost with measured evidence

Why it matters

Cost optimizations only count when measured: a cheaper route that retries or a cache that serves stale answers can cost more than it saves.

What to practise

Run synthetic requests against live Microsoft Foundry deployments as a premium control, then under an optimized policy. You implement a router that assigns tiers without spending model tokens, a context budget enforced before the call, and Azure Managed Redis keys for exact results and built prompt context that include agent and policy versions. A quality floor moves a weak answer up a tier and flags human review only at the top tier. Invalidate one exact result, confirm the next run falls back to the prompt-context level, and compare the runs’ evidence rows. All three tiers map to one deployment, so the differences come from budgets, retries and caching.

How it fits

Versioned keys and targeted invalidation are 2.4.8’s freshness scope; the optional challenge adds its tenant scope. The prompt-context level is an application cache, not 2.4.7’s provider prompt caching. For 3.2, note which cheaper routes retried and what that cost.