Design-agent memory system raises execution success in preprint tests

Researchers report that an external library of reusable procedures raised a Claude Sonnet 4 design agent's execution success from 72.7% to 99.3% across five adaptation rounds. The preprint's Designer-RSI system operated professional design software through more than 230 tools while keeping the underlying model frozen. The system expanded its procedure library from 76 to 139 skills using 1,406 user briefs and 1,869 automatically graded trajectories. A replay gate retained revisions that repaired failures without undoing prior successes. Results rely on automated grading and benchmark comparisons rather than human labels or production deployment.