May 2025

Examines agents that execute code, retain external memory and revise their actions. Discusses tool isolation, execution controls and training with verifiable rewards; it is a feasibility study, not a performance result.