MLflow 3.16 made the redesigned trace explorer default and patched basic-auth security.

MLflow released version 3.16.0, making the redesigned trace explorer its default interface [16]. The release added session grouping for multi-turn conversations, span links, and per-user budget policies in the AI Gateway [16]. MLflow 3.16.1 followed shortly after to patch a hardcoded administrative password in basic authentication, while also adding span link support for Unity Catalog traces and resolving trace location lookups in Model Serving [11]. In parallel, MLflow 2.11.5 introduced an option to route Unity Catalog model registry artifact transfers through the Databricks SDK Files API [7].

LLM evaluation and agent tracking remained central throughout the month. Production case studies demonstrated evaluation-first architectures that paired traces with automated judges to maintain quality gates [13], while practitioners documented patterns for tracking per-run agent costs with LangGraph [10]. Other evaluations tested whether specialized evaluation models like Jev could cut latency and compute spend for MLflow question-answering assessments relative to general-purpose frontier models [4] [8].

Community members faced operational hurdles around experiment management and trace persistence. Practitioners worked through behavioral differences between Workspace and Unity Catalog experiment targets when using autologging [17]. At the end of the month, users reported an issue where StartTraceV3 returned success codes but left traces unretrievable via GetTrace [1].