LoginSign Up
← Back to ai-ml-platforms
Weekly · ai-ml-platforms · Week of October 19, 2026

RAGFlow, LiteLLM and MLflow ship agent features and the controls to contain them in the same releases

agentsleast-privilegeobservabilitymcpgo-rewrite
O
Generated 1h agoDrawn from 6 products

The week in ai-ml-platforms

The center of gravity in AI and ML platforms is agents, and the tooling around them is hardening. RAGFlow used its second 1.0 release candidate to replace chat thinking levels with a single Agentic mode, LiteLLM opened 1.107 with a default-deny switch for agent access, and MLflow is recasting its UI around GenAI traces. Capability and control arrived in the same releases.

A second pattern is observability. Arize Phoenix started exporting its own server logs over OTLP, and MLflow made judges return typed decisions. Evaluating and tracing agents is now a core part of the platform, not an add-on.

Leaders

RAGFlow rewrote itself in Go for v1.0.0-rc1 at the end of September, moving messaging from Redis to NATS and cache and checkpoints to Kvrocks. RC2 then replaced the Low, Medium, High and Ultra thinking levels with an Agentic chat mode, unified parsers and added arm64 images, and restored Team and Me sharing that RC1 dropped.

LiteLLM took 1.106.0 stable after four release candidates, with stdio MCP servers off by default and a one-command Docker quickstart that starts the gateway, Postgres and admin UI. The 1.107 release candidate adds a least-privilege agentaccessdefault_deny switch and verified stateless MCP.

Arize Phoenix released 20.22.0 with new playground models, an OAuth cookie fix and Claude Managed Agents tracing docs, then 20.23.0 a day later, which lets the server ship its own logs over OTLP and fixes Bedrock tool results in the playground.

MLflow is centered on LLM tracing and evaluation rather than experiment tracking. 3.16 rebuilt the trace explorer so the Assistant can generate custom trace views from plain English, and 3.17 added typed judge feedback and resource-level permissions. 3.17.1 fixes Claude Code trace rendering.

Wildcards

AnythingLLM spent two straight releases on fixes, many from outside contributors, after a feature run that took dictation, text actions and autocomplete across the desktop and added a collapsible activity view for long agent runs.

Kubeflow published only a redirect notice: the main repository no longer holds project code, and releases now come from seven subproject repositories, each on its own cycle.

Themes that compounded

  • Agent modes replacing manual knobs: Agentic chat at RAGFlow.
  • Least-privilege defaults for agents and MCP at LiteLLM.
  • OpenTelemetry-style logging and tracing at Arize Phoenix and MLflow.
  • Security-relevant breaking changes: MLflow 3.16.1 no longer ships a default basic-auth admin password.

Watch this week

Watch LiteLLM's 1.107 release candidates for a stable tag and for the first operators reporting on default-deny agent access. RAGFlow's 1.0 line should reach a third candidate as Go ingestion fills gaps, and Phoenix's two-day patch cadence is worth tracking alongside MLflow's 3.17 series.