“AI is slow” is not a diagnosis. Separate sensing, decision selection, spatial queries, paths, movement and animation integration.
Choose rates by urgency
Collision response may run frequently; strategic goal selection can run less often or on events. Stagger periodic updates across frames.
Budget asynchronous work
Path and planning queues need latency and throughput targets. Define priority, cancellation and agent behaviour while waiting.
Control fan-out
Perception pairs and candidates grow with agent count. Use spatial partitioning, relevance filters and early rejection. Cache only with explicit assumptions.
Profile quality
Track CPU with reaction latency, stuck time and decision churn on shipping hardware. The archive’s Game AI vs. Traditional AI explains why more CPU alone is not the answer.