MLflow 3.15 introduced immutable evaluation datasets, ensemble scoring, and serverless runtime fixes.

In August 2026, MLflow received two patch updates focused on evaluation workflows and runtime reliability. MLflow 3.15.1 corrected version parsing issues on Databricks Serverless and adjusted packaging behavior on ARM client images [12]. MLflow 3.15.2 introduced immutable evaluation dataset versions and the scorer_ensemble primitive for aggregating judge outputs [2]. The patch also resolved a Databricks telemetry deadlock and fixed edge cases in aligned judge execution and run status constraints [2].

Databricks tooling expanded its integration with MLflow storage and tracking. Both the Databricks SDK and platform job definitions gained support for configuring custom MLflow artifact locations within automated tasks [7, 8]. In addition, the Databricks AI Runtime command-line tool surfaced native MLflow experiment tracking alongside hardware utilization metrics for distributed LLM training runs across serverless GPU clusters [4].

AI agent observability and evaluation remained prominent themes throughout the month. Educational materials detailed how to capture agent execution steps with MLflow traces when orchestrating multi-tool systems [9], paired with guidance on evaluating agent skills [13]. Meanwhile, community discussions reflected practical demands from teams seeking established step-by-step blueprints to operationalize language model evaluations in production Databricks environments [1].