FDE Times: build an agent that asks when a request is ambiguous — then measure whether asking pays off
A 6 November FDE Times guide shows FDEs how to wire AskUserQuestion into the Claude Agent SDK, score under-asking and over-asking separately, and compare against a guessing baseline so clarifying questions only stay if they improve results.
Customer requests at a site almost always lack context. FDE Times walks through a Claude Agent SDK agent that asks via AskUserQuestion, with a hard cap on rounds, then an eval that measures whether asking is worth the interruptions.
Prompt tricks alone are weak: CLAMBER and related benchmarks show models still under-ask on ambiguous inputs and write poor clarifying questions. The guide treats “when to ask” and “what to ask” as separate skills, keeps the tool on the main agent (not subagents), and forces concrete options rather than vague “what do you want?” prompts.
The decisive step is a side-by-side run on labelled cases — with and without the ask tool — reporting under-asking rate, over-asking rate, pass rate, and user interruptions. Asking only earns its keep when the ambiguous group’s passes rise and clear requests stay untouched; otherwise drop it.
For FDEs, that comparison table is what you bring to a customer review instead of claiming the agent “now knows how to clarify.”
Source: fdetimes.net