The tweet was deleted by the author.
But we saved everything 🙂.
Databricks has introduced the coSTAR pattern for building software agents, according to Matei Zaharia. The approach is designed to improve agents’ ability to address users’ most challenging tasks while avoiding regression in their performance. Zaharia notes that the effectiveness of coSTAR is amplified when integrated with more advanced models and that its implementation is streamlined in MLflow, Databricks' open-source platform.
Zaharia has previously spotlighted Databricks Research with Harvard and Cornell on off-policy reinforcement learning outperforming on-policy methods. He also reported that DSPy is now supported on MLflow and Databricks using their evaluation interfaces. These developments reflect ongoing integration of research and tools within the Databricks ecosystem.