Professional graphic design is a long-horizon agentic task: structured, editable artifacts emerge from many interdependent actions, yet the outcome offers no reliable programmatic oracle. A new Hugging Face paper, Designer-RSI: Evolving Procedural Memory from User Traffic for Agentic Graphic Design, proposes a continual-adaptation framework that sidesteps this problem by leaving the underlying model untouched.
In the setup, a frozen frontier model operates professional design software through more than 230 tools. What changes over time is an external procedural memory: a bank of natural-language skills that accumulates and refines reusable design procedures from experience. The memory widens by acquiring procedures for recurring subtasks that no existing skill covers, and deepens by revising existing procedures against their own successful and failed executions.
A matched replay gate governs every update: a candidate change is admitted only if it repairs at least one replayed failure without regressing observed successes. Everything else — the models, tools, and grader — stays frozen, and no human labels are used.
Across five rounds over 1,406 real user briefs and 1,869 automatically graded trajectories, the skill bank grew from 76 documentation-derived skills to 139. GenEval2 execution success on Claude-Sonnet-4 rose from 72.7% to 99.3%, with a gain of 11.99 points in generation quality. Against a no-skill agent, win rates reached 61.8% on Claude-Sonnet-4 and 67.6% on Claude-Opus-4.6 across four specialized design benchmarks.
The authors also isolate the contribution of each mechanism. On 200 held-out briefs from a user-traffic benchmark, widening alone reached a 49.4% win rate and deepening alone 48.6%, while combining both reached 58.5% (p = 0.025). The takeaway: when outcomes cannot be checked automatically, an external procedural memory offers a practical route to continual adaptation for frozen agents — and every improvement is a skill edit a person can read and audit.