Paper 1 · Question 27

A: Better descriptions help but do not solve the fundamental problem: 18 tools exceed the recommended range and create too much decision complexity. Selection accuracy degrades regardless of description quality when tool count is too high.

B: A routing classifier adds infrastructure complexity and bypasses the LLM's natural language understanding. The root cause is too many tools, not inadequate routing.

C: Forcing tool selection does not help when the agent cannot reliably choose among 18 options. The problem is selection accuracy due to overload, not whether the model calls a tool at all.

D: The recommended range is 4-5 tools per agent. Splitting into role-specific agents (for example separate agents for code analysis, security scanning, and test generation) eliminates decision complexity and prevents cross- specialisation misuse. Each agent excels within its narrow remit.