Most coding-agent benchmarks hand the agent a fully specified task, a bug report with a clear fix, or a feature spec with unambiguous requirements, and grade whether it gets that fixed target right. ICAE-Bench argues that's no longer how people actually use coding agents in the vibe-coding era, where someone describes a rough product idea and expects the agent to clarify requirements, make design decisions, and build the thing through back-and-forth, much like working with a junior engineer rather than issuing a spec. So the benchmark simulates that dynamic directly: each task starts from a real open-source repository with genuinely fuzzy, underspecified intent, and an automated User Agent plays the role of the human, answering the coding agent's clarifying questions and revealing hidden requirements over the course of the interaction, but only ever revealing real constraints pulled from ground-truth data, never inventing new ones or leaking implementation details that would let the coding agent cheat. To make this fair to grade despite the open-ended nature of the task, the authors combine standardized black-box tests with a multi-dimensional scoring rubric covering functional correctness, semantic and API similarity to the reference implementation, structural fidelity, design quality, and even the quality of the back-and-forth interaction itself. That's a meaningfully different bar than does the code pass the tests, and it's directly useful for anyone building or evaluating an agentic coding product: if your agent is good at executing precise instructions but bad at asking the right clarifying questions or making sane design calls when requirements are vague, static benchmarks won't catch that gap, ICAE-Bench is built specifically to expose it. For teams building internal coding agents or picking between commercial ones, this kind of interactive, requirements-discovery evaluation is a better proxy for real build-me-a-feature usage than most existing static benchmarks.