Five agents asking one question should cost one answer
Multi-agent systems fan out, and they ask overlapping questions constantly. Without a shared cache, that overlap is pure waste, the same answer, bought once per agent.
The moment you go multi-agent, a new cost appears that single-agent apps never see: overlap. Five agents working a problem will independently ask for the same schema, the same policy, the same definition, phrased a little differently each time. Without something in the middle, that's five model calls for one piece of knowledge.
The first agent pays; the rest reuse the answer for free.
As you add agents, this only gets better. More agents means more overlap, and more overlap means a higher cache hit rate. The architecture that scales your costs the fastest, fan-out, is exactly the one a shared semantic cache tames the best.
The bottom line
In a multi-agent system, shared knowledge should be a shared cost. One answer, reused across the swarm, instead of one bill per agent.