Professional graphic design is a long‑horizon agentic task where structured, editable artifacts emerge from many interdependent actions, yet outcomes lack a reliable programmatic oracle. We introduce a continual‑adaptation framework: a frozen frontier model operates professional design software through more than 230 tools, while an external procedural memory of natural‑language skills accumulates and refines reusable design procedures from experience. The memory expands via two mechanisms:
- Widening: acquiring new procedures for recurring uncovered subtasks;
- Deepening: revising existing procedures against their own successful and failed executions. A matched replay gate admits only changes that repair failures without regressing observed successes.
Across five rounds covering 1,406 real user briefs and 1,869 automatically graded trajectories—without any weight updates or human labels—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%, a 11.99‑point gain in generation quality. Against a no‑skill agent, win rates on four specialized design benchmarks were 61.8% (Claude‑Sonnet‑4) and 67.6% (Claude‑Opus‑4.6).
On a held‑out set of 200 briefs from a user‑traffic benchmark, widening alone achieved a 49.4% win rate, deepening 48.6%, while their combination reached 58.5% (p = 0.025). Procedural memory thus offers a practical route to continual adaptation of agents under noisy, unverifiable feedback.
Review: This work demonstrates that an external, editable procedural memory can enable zero‑weight‑update adaptation of large models in complex design tasks, balancing interpretability and scalability, and paving the way for more interactive creative AI systems.