Designer-RSI: Evolving Procedural Memory from User Traffic for Agentic Graphic Design
A frozen design agent jumped from 72.7% to 99.3% execution success by growing and revising an external skill memory.
The paper describes a graphic-design agent using Claude-Sonnet-4 with no weight updates, operating professional design software through more than 230 tools. Across five rounds on real user briefs and automatically graded trajectories, its procedural memory grew from 76 documentation-derived skills to 139. The authors say the system improves by adding procedures for missing subtasks and revising existing ones, with a replay gate meant to avoid changes that fix one failure while breaking prior successes. HF Daily Papers' note
The paper describes a graphic-design agent using Claude-Sonnet-4 with no weight updates, operating professional design software through more than 230 tools. Across five rounds on real user briefs and automatically graded trajectories, its procedural memory grew from 76 documentation-derived skills to 139. The authors say the system improves by adding procedures for missing subtasks and revising existing ones, with a replay gate meant to avoid changes that fix one failure while breaking prior successes. HF Daily Papers' note
score 5