OpenAI-Compatible Agents with Governed Context: Feast, MLflow, and OGX

Most agent stacks stop at retrieval: pull context, call the model, return an answer. In production, that is not enough. You still need to know what context the agent saw, whether the response was good, and how to turn real interactions into training data without rebuilding the pipeline every time.

In this talk, we show that loop end to end with Feast, MLflow, and OGX(Llama Stack). An OpenAI-compatible client calls OGX’s Responses API - server-side agentic loop, MCP tools, and built-in RAG - while Feast serves structured features and vector context through its OpenAI-compatible retrieval APIs. MLflow traces each run and records which Feast entities and features were served. Human reviewers label interactions via Feast feedback writes; Feast then joins MLflow traces with those labels into a fine-tuning dataset, with lineage through deployment.

This is not a framework bake-off. It is a proof that OpenAI-compatible agent serving, governed context, and observability can share one workflow: OGX for generation and tools, Feast for context and feedback, MLflow for traces and dataset assembly.