Claude Certified Architect – Foundations Practice Question 69: Tool Search Deferred Loading
Claude Certified Architect, Foundations · Test 7 · question 9 of 10
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?
Show answer and explanation
Answer:
A is correct. The tool search tool lets Claude discover tools on demand. You include a tool search tool, send every definition in the tools array, and mark the tools that should not load up front with defer_loading: true. The docs say Claude's tool selection degrades past roughly 30-50 tools and that tool search typically cuts definition tokens by over 85 percent. Deferred tools are excluded from the system-prompt prefix, so prompt caching is preserved, and keeping the 3-5 most used tools non-deferred is recommended.
B is wrong because defer_loading controls what enters the context window, not what you send: you must still send each tool's full definition, which the API needs server-side to run the search. C is wrong because at least one tool, normally the tool search tool, must stay non-deferred; deferring everything returns a 400 error. D is wrong because a larger window does not fix the accuracy problem or the per-request cost of 60k tokens of definitions.
Reference: Tool search tool - Claude API Docs