SCLATE: A Substrate for Continual-Learning Agent Training and Evaluation
Apple just open-sourced the infrastructure layer for training continual-learning agents at scale — solving the scheduling mess every agent team rebuilds.
Why it matters
SCLATE addresses a concrete operational gap: agent benchmarks and frameworks each build custom event schedulers for multi-session, memory-aware agent training. A shared substrate with hybrid simulated clocks lets benchmarks and unmodified agents plug into one scheduler, reducing custom integration work and enabling standardized agent evaluation across heterogeneous setups.
The key facts
14 to knowPublished by Apple Machine Learning Research
SCLATE: execution substrate with open event scheduler
Solves interleaving of benchmark events and agent-side events (session start/stop, crons, memory consolidation)
Adapters allow benchmarks and agents to add events without modification
Hybrid simulated clock architecture (specifics not disclosed in abstract)
Addresses the 'each benchmark and agent pair builds custom scheduling loop' problem
No performance metrics, deployment data, or availability timeline disclosed
SCLATE: open event scheduler adapter for benchmarks and agents
Solves scheduling of multi-session agent evaluation with memory consolidation events
Each benchmark and agent adds events through adapter; hybrid simulated clock coordinates timing
Eliminates custom scheduling loop per benchmark-agent pair
Published by Apple Machine Learning Research, Sep 30 2026
No adoption data, pricing, or integration details disclosed
Designed for continual-learning agents operating over long horizons
Go to the source
Apple Machine Learningmachinelearning.apple.com
Publisher excerpt: Continual-learning agents are systems of models, harnesses, and memory operating over long multi-session horizons. Evaluating and training them requires interleaving tasks with agent-side events such as session stop and start, crons, and memory consolidation. Yet existing benchmarks and training…