SCLATE: A Substrate for Continual-Learning Agent Training and Evaluation

SCLATE is a new execution substrate enabling unified training and evaluation of continual-learning agents by integrating benchmarks and agent events on a shared timeline, compressing long scenarios into hours. It supports unmodified agent components, facilitating head-to-head comparisons and post-training improvements, demonstrated with Qwen3.5-4B's enhanced memory use and task performance.