Tag
The paper introduces JAZ, a minimalist agent framework that uses a single invoke primitive to enable LLM agents to handle long-horizon tasks and self-improvement without external harnesses, outperforming existing systems like MemGPT and ACE in evaluations.