The Case for Minimalist Agent Architectures

Modern agent frameworks often rely on complex, specialized systems for memory management, file handling, and self-improvement. The authors argue that these external harnesses are often unnecessary. They introduce JAZ, a framework that reduces agent architecture to a single, recursive primitive called invoke. By treating the agent loop as a language primitive rather than an external wrapper, JAZ achieves high expressivity with minimal overhead.

The 'Invoke' Primitive

JAZ is built on two core principles that define its invoke primitive:

  1. Recursive Execution: The LLM is empowered to write arbitrary executable code that can call invoke recursively, allowing the model to manage its own control flow and state transitions.
  2. Environment Transparency: All inputs to invoke and the entire interaction history are exposed as variables within the code environment. This allows the LLM to treat its own history and context as manipulatable data.

By viewing invoke as a function whose implementation is provided by the LLM at runtime, the framework eliminates the need for manually designed tools or rigid external systems. The system provides basic hooks for monitoring and constraints, but the logic remains entirely within the LLM's generated code.

Performance and Efficiency

To validate this minimalist approach, the authors tested JAZ against specialized systems without using any pre-built memory or file system tools. The results demonstrate that a simple, code-centric loop can outperform purpose-built frameworks:

  • Long-Horizon Recall: On the StuLife benchmark, JAZ outperformed Letta (formerly MemGPT) by 8% while operating at half the cost.
  • Continual Self-Improvement: On the AppWorld benchmark, JAZ outperformed ACE by 4% at a lower cost.

These results suggest that the current trend of building increasingly complex agent harnesses may be less effective than simply providing LLMs with a clean, recursive interface to interact with their own execution environment.