[Meta Releases CORAL Framework to Optimize Recommendation Systems with LLM Agents] DeepFlow Tech News, September 3, Meta released the paper CORAL, introducing a native LLM agent harness designed for closed-loop optimization of production-grade recommendation systems. This system continuously reallocates online recommendation strategies within a fixed change budget by observing operational signals in a closed loop, recalling past decision memories, and utilizing tools like numerical optimizers. The paper demonstrates that strategies can gradually improve through iterative cycles without requiring parameter updates.