Most existing coding-agent benchmarks measure something narrower than how these tools actually get used day to day: given a specific, well-defined bug and a failing test, can the agent produce a patch that makes the test pass. That's a useful and tractable thing to measure, but it doesn't capture what a lot of real-world usage actually looks like, which is starting from a vague, incomplete idea of what someone wants built and having to figure out the rest along the way. ICAE-Bench is built around that gap directly: instead of scoring agents on bug-fixing accuracy against a pre-defined target, it evaluates how well an agent can take an underspecified product intent, ask the right clarifying questions or make reasonable assumptions where the request is ambiguous, plan out an approach, and then actually construct a working piece of software at the repository level rather than a single isolated function or patch. This shift matters because it tests a fundamentally different and arguably harder skill: judgment under ambiguity, rather than pure execution against a clear specification. A model can be excellent at turning a precisely defined bug report into a correct fix while still being weak at recognizing that a vague request is missing important details, or at making sensible architectural decisions when nobody has specified exactly how a feature should be structured. For teams evaluating which coding agent to adopt for genuinely open-ended tasks, like "build me a tool that does X" rather than "fix this specific failing test," a benchmark like this is a more honest predictor of real performance than the bug-patching benchmarks most current agents are optimized and marketed against.