JAZ’s minimalist design for self-improving AI agents
A user highlights an MIT CSAIL paper describing JAZ, a framework in which a language model writes code and can recursively call a single operation, invoke.
TLDR
According to a user’s account of an MIT CSAIL paper, JAZ gives the model access to its inputs and history as variables in its code environment. Memory and self-improvement are handled by code the agent writes inside the loop, with hooks for constraints and monitoring.
The user says JAZ, using prompting alone and no memory system, beats Letta (MemGPT) by 8% at half the cost on the recall-heavy part of StuLife. They also report that it beats ACE, a self-improvement method, by 4% at lower cost on AppWorld.
