Apple Machine Learning researchers have introduced SCLATE, an execution substrate designed to streamline the training and evaluation of continual-learning agents—systems that operate over long, multi-session horizons.
Product LaunchesAppleApple Machine Learning
Apple Researchers Introduce SCLATE Substrate for Continual-Learning Agent Training and Evaluation
Existing benchmarks and training frameworks often require custom scheduling loops for every benchmark-agent pair. SCLATE addresses this by providing a shared execution environment where benchmarks and unmodified agents add events to a single open event scheduler through an adapter. The system utilizes a hybrid simulated clock to run these events on a shared timeline, allowing a month-long scenario to be compressed into a few hours.
The platform also functions as a rollout engine, recording token data and log probabilities through an in-container proxy without requiring modifications to the agent's harness or memory. In tests, researchers ported seven benchmarks to SCLATE and compared ten unmodified harness and memory configurations across ten different models. The results indicated that adding a separate memory system does not reliably outperform a harness's native memory, and that models exhibit significant variance in how they utilize the same harness and memory.
Furthermore, the researchers applied post-training to Qwen3.5-4B using unmodified harnesses and memory systems. The study found that the model improved its efficiency and performance, reading 6.8× fewer file lines while achieving a 16.7-point higher pass rate on SWE-bench Verified and showing increased accuracy on MetaClaw.
Sources
- SCLATE: A Substrate for Continual-Learning Agent Training and Evaluation （Apple Machine Learning, 2026-09-30）
- arXiv