MLflow moves from grading agents before release to watching them in production, and its neighbours track runs and breaking changes.

The week in ai-ml-platforms

Only four products in this sector were updated, so this is a short report. The lead is MLflow 3.18, which treats agent evaluation as something that continues after release.

The other three are about repeatability and plumbing: comparing runs, migrating routers and improving document parsing.

Leaders

MLflow released 3.18.0, which clusters recurring agent failures into investigable issues, scores live traffic with code-based checks as traces arrive, and adds a registry for sharing agent skills. The earlier 3.17 added typed judge feedback and resource-level permissions with explicit DENY.

Arize Phoenix shipped Python client 3.7.0, which saves model parameters and agent settings with experiments so two runs can be compared on what changed, and makes annotation uploads fail early when key columns are bad. Client 3.6.0 added agent annotations and trace filters.

LiteLLM published 1.109 dev.2 with breaking changes to customer pagination, router cooldowns and fallback rules, and per-model quota requirements for projects. The 1.109 release candidate priced PDF pages per provider billing.

RAGFlow nightlies added a Somark OCR connection check and image-table recognition in DeepDoc, as the project moves toward 1.0.

Themes that compounded

• Monitoring and scoring live agent traffic at MLflow.
• Run comparability at Arize Phoenix through saved settings.
• Migration lists for operators at LiteLLM.
• Parsing quality at RAGFlow.

Watch this week

LiteLLM's 1.109 line is in dev builds with a breaking list, so a release candidate with migration notes is the next item. MLflow's agent skills registry is new, and RAGFlow's nightlies point toward 1.0.