Agents Flunk the Real Design Grind

🔥 What happened Researchers pitted Codex/Astra against Claude Code on 35 design briefs across 43 conditions. Astra cleared all 43; Claude managed 35. On 28 shared core tasks, semantic checks flagged 11 violations for Astra and 5 for Claude. 💡 Why it matters Structural credit was perfect across the board — the semantic layer is where it cracks. That gap between execution, requirement compliance, and real-world acceptance is exactly where agent readiness gets decided. ⚡ Our take A benchmark that separates execution from compliance is overdue. Grading agents on "it runs" measures the wrong thing.