Cutting agent latency from 30s to 8s without model swap

An AI chat team reduced p95 latency from 31s to 8s without changing models.

Only about 35% of the delay came from the model; most was caused by sequential tool calls, extra LLM steps, and no streaming.

They parallelized independent lookups, replaced an LLM critic with deterministic validation, and streamed responses to users.

The result was a 70% drop in abandonment, showing orchestration often matters more than model speed.