Paper proposes JAZ framework using a single LLM primitive to replace specialized agent harnesses
JAZ’s 'invoke' primitive unifies memory and self-improvement workflows without external systems, outperforming MemGPT and ACE baselines on long-horizon recall and continual self-improvement tasks.
1 source · cross-referenced
- JAZ introduces a minimal agent framework centered on a single LLM primitive called 'invoke' that generalizes code-mode agent loops.
- The framework eliminates the need for external memory or self-improvement systems by exposing all inputs and history as variables in the code environment.
- On long-horizon recall tasks, JAZ’s invoke outperformed Letta (MemGPT) by 8% at half the cost on the recall-heavy portion of StuLife.
- On continual self-improvement tasks, JAZ’s invoke outperformed ACE by 4% at a lower cost on AppWorld.
Researchers propose JAZ, a minimalist agent framework built around a single LLM primitive called 'invoke' that generalizes code-mode agent loops. The primitive allows the LLM to write arbitrary executable code, including recursive calls to invoke, while exposing all inputs and interaction history as variables within the same code environment. This design eliminates the need for external memory systems or self-improvement harnesses by treating the agent loop itself as the primary abstraction.
The paper motivates the design from first principles, framing invoke as a language-level function whose implementation is provided at runtime by the LLM. The framework provides built-in hooks for programmers to apply constraints and monitoring, but does not rely on manually designed tools or external systems to achieve its reported results.
To validate the approach, the authors evaluate invoke on workflows traditionally requiring specialized external systems. On long-horizon recall tasks that exceed typical context windows, JAZ’s invoke outperformed Letta (MemGPT) by 8% while using half the cost on the recall-heavy portion of the StuLife benchmark. On continual self-improvement tasks, JAZ’s invoke outperformed ACE by 4% at a lower cost on the AppWorld benchmark.
The authors note that their evaluation used only prompting and did not incorporate manually designed tools, external harnesses, or systems for memory or file system access, relying instead on the core invoke primitive and its built-in hooks.
- Nov 11, 2026 · Latent Space — swyx
OpenAI demonstrates voice steering, cloud handoffs, and personal agents for Codex at AI Engineer Code Summit
Trust79 - Nov 11, 2026 · The Verge — AI
OpenAI to launch verified AI Business Agents in ChatGPT for US e-commerce and service businesses
Trust79 - Nov 10, 2026 · Latent Space — swyx
NaiveAI releases Naive-N0.5-Flash coding model with ultrafast token throughput
Trust74