Professional design work involves understanding a client's vague request, exploring creative options, making edits based on feedback, and preparing final files for production. Current AI agents can carry out these steps, but it is unclear if they truly understand what was asked for.
To test this, researchers created DRAFT-Bench, a set of 35 design assignments with 43 different conditions. They used it to compare two AI systems: Astra and Claude Code, scoring how many tasks they completed and whether the results actually met the original requirements.
Astra completed all 43 conditions and Claude completed 35 out of 43. However, out of 28 core tasks where both systems were scored, semantic checks found requirement violations in 11 for Astra and 5 for Claude. This shows completion does not guarantee compliance, separating execution from true professional acceptance.