Researchers proposed JAZ, a minimal agent framework. It uses a single primitive called invoke. Invoke lets the LLM write and run code that can call itself. Everything the LLM sees is a variable in that code. JAZ was tested on recall-heavy and self-improving tasks. It beat Letta by 8% on recall tasks at half the cost. It beat ACE by 4% on self-improvement at lower cost. The paper argues invoke can replace specialized harnesses. It does this without extra tools or memory systems. The work was published on arXiv.

• 8% higher recall on StuLife
• 4% better on AppWorld
• half the cost of Letta
• lower cost than ACE