CCARF Practice Question

Paper 2 · Question 2

D2 Tool Design & MCP Integration · Paper 2

The data platform team has built an MCP server with 22 tools: query_snowflake, query_postgres, query_api, plus 19 specialised tools for individual data transformations (pivot_table, calculate_percentile, normalise_currency, etc.). Agents take 3-4 turns to select the correct tool and frequently choose the wrong transformation. What is the most effective redesign?

• A.Improve all 22 tool descriptions with detailed examples and boundary conditions to help the agent distinguish between them.
• B.Consolidate the 19 transformation tools into a single transform_data tool with a transform_type parameter, reducing the total to 4 tools.
• C.Use tool_choice: 'any' to force the agent to always call a tool, eliminating turns where the agent reasons without acting.
• D.Split the tools across two separate MCP servers — one for queries and one for transformations — to reduce cognitive load.

Show answer & explanations

Correct answer: B

A: Better descriptions help, but 22 tools still exceeds the practical limit for reliable selection. Research shows tool selection degrades significantly beyond 4-5 tools per agent. The tool count itself is the problem.

B: Reducing from 22 to 4 tools brings the count within the optimal 4-5 range for reliable tool selection. The 19 transformation tools share a common pattern (input data, transformation type, output format) and are natural candidates for consolidation into a single parameterised tool.

C: Forcing tool calls does not improve selection accuracy — it just ensures the agent picks something, potentially the wrong tool. The root cause is too many similar tools, not too few tool calls.

D: MCP server boundaries are invisible to the agent. All tools from all connected servers appear in a single list. Splitting across servers does not reduce the number of tools the agent must choose from.

Related Tool Design & MCP Integration questions

More: Paper 2 · All practice questions