Harness as a Language (academy.dair.ai)

A new framework called JAZ has been introduced, designed to enhance language-model agents by simplifying their operational structure while maximizing their capability to tackle complex workflows. Unlike existing systems that rely on intricate architectures with external tools and memory management, JAZ centers around a minimalist agent loop featuring a single LLM-based primitive called "invoke." This innovative approach allows the LLM to generate executable code, including recursive calls to the "invoke" function, enabling it to interact with its environment dynamically while maintaining a comprehensive view of its execution history. The significance of JAZ lies in its ability to effectively manage long-horizon workflows and self-improvement tasks without the need for additional engineering resources, outperforming established systems such as Letta (MemGPT) and ACE in specific evaluations. In tests involving recall-heavy tasks on the StuLife dataset, JAZ's "invoke" demonstrated an 8% improvement at half the cost, while it also achieved a 4% performance edge over ACE on continual self-improvement tasks within AppWorld. This promising approach suggests a new direction for developing efficient and capable LLM agents, potentially transforming how such systems operate within the AI/ML community.