An enterprise agent aggregates five MCP servers and exposes about 180 tools to the Claude API. The definitions use roughly 60k tokens on every request, Claude increasingly picks the wrong tool, and the team wants to keep prompt caching working. Which design is correct?