Apple researchers released a framework for testing AI agents

Researchers from Apple introduced SCLATE, a new platform designed to help AI agents learn and perform better over long periods. Currently, many agents struggle to maintain their capabilities across multiple sessions because they start each task from scratch.

SCLATE acts as a central scheduler for these systems. Instead of forcing developers to build custom loops for every project, it allows different benchmarks and agent setups to plug into a shared timeline. It runs these tasks in real-time but skips periods of inactivity, meaning a process that would normally take a month can be completed in hours.

This framework also functions as a testing engine. It records exactly how a model makes decisions by tracking its internal tokens and logs. By using SCLATE, the research team compared ten different ways to manage memory and task harnesses across several models.

The findings show that adding extra memory systems does not always improve performance over the built-in methods. However, when the team trained the Qwen3.5-4B model using these standardized harnesses, it became significantly more efficient.

After training, the model read nearly 7 times fewer lines of code while achieving a higher success rate on complex coding tasks. It also wrote more detailed notes to its own memory, which improved its performance on held-out tests.

How much better would your own assistant perform if it could actually remember your preferences from a year ago?

Despite these improvements, SCLATE currently relies on specific test benchmarks ported to its system. It does not yet account for every possible real-world variable, and the study did not test all existing agent architectures on the platform.