Weekly ML Roundup: Agentic RAG in SQL and Durable Agent Memory

by TechHub

This week in ML, the focus shifted toward running agent workflows closer to the data tier, with SQL-driven agentic RAG and chat completion patterns that bring prompts, tool calls, and logging under existing database security and auditing. We also saw practical guidance on agent isolation and durable memory using Azure AI Foundry with Azure Blob Storage, plus production-oriented debugging considerations when agents run as long-lived services. On the applied side, Microsoft Discovery highlighted a human-in-the-loop, multi-agent approach to wastewater metagenomics that keeps reproducibility and reviewability central. Rounding out the week, platform updates covered the operational plumbing that hybrid analytics and agent-backed pipelines depend on, including gateway improvements and expanded observability for Oracle AI Database@Azure.

This Week's Overview

Durable AI agents and agentic RAG inside the data tier

A common thread this week was pushing agent workflows closer to where data already lives, while making agent state more reliable and auditable, picking up on last week's emphasis on governed grounding (lineage, catalog search, and traceability) by showing what it looks like when the “agent loop” runs directly in the database alongside existing security and data logic. Microsoft content showed patterns for agentic RAG (retrieval-augmented generation) and chat completion running “in the SQL engine” via REST calls to model endpoints, and practical approaches for isolating agents and giving them durable memory using Azure storage.

Agentic RAG and chat completion driven from SQL

The “Data Exposed” walkthrough shows SQL Server, Azure SQL, and SQL DB in Fabric acting as an orchestration layer for agentic RAG and chat completion, extending last week's “make governance operational” theme by moving prompt execution and logging into a tier that many teams already treat as the source of truth for permissions, auditing, and change control. The key idea is straightforward: T-SQL invokes REST calls to model endpoints, SQL processes the returned payloads, and the database becomes the place where prompts, tool calls, and results can be integrated with existing data logic and governance.

For teams already standardizing on Azure AI Foundry, the demo connects the pieces end-to-end, including Foundry Gateway as the model access layer and Microsoft Teams as a user-facing entry point. This style of architecture can reduce the amount of application glue code you need for basic RAG patterns, but it increases the importance of prompt safety in the data tier, including defenses against prompt injection when your retrieval content comes from untrusted sources.

Agent isolation and durable memory with Foundry and Blob Storage

Episode 7 of “The Upload” adds a pragmatic angle to agent building: if your agent needs memory beyond a single interaction, you need a durable store and a clear isolation model, which mirrors last week's case studies where “shippable” agents depended on durable history plus audit-friendly state. The segment calls out using Microsoft Foundry with Azure Blob Storage to persist agent memory, which is often the difference between a prototype chatbot and an agent that can resume work after restarts, failures, or deployments.

The same episode pairs this with production-oriented debugging topics (like .NET memory dumps) that matter when you start running agents as always-on services. The takeaway for developers is to treat memory and state as first-class parts of the design: define what is persisted, how it is keyed per user or task, and how you will inspect it when things go wrong.

Microsoft Discovery and AI agents for scientific metagenomics workflows

These posts focus on a more “human-in-the-loop” agent pattern: using multiple collaborating agents to accelerate exploratory analysis, while keeping scientific judgment and reproducibility at the center, continuing last week's Microsoft Discovery storyline by applying the same “durable memory plus provenance” idea to a new scientific workflow. The example domain is wastewater metagenomics, where teams need to filter and interpret large sequencing datasets and relate findings to environmental context like local weather.

Wastewater virome analysis with orchestrated agents and dashboards

Microsoft describes using the Microsoft Discovery app to orchestrate AI agents across a workflow that includes taxonomic filtering, weather-data enrichment, phylogenetics, and interactive dashboards, building on last week's closed-loop Discovery case study by showing another concrete pattern for keeping agent-assisted science reviewable and reproducible. The emphasis is on amplifying expert work rather than replacing it, using rubric-driven checks and explicit decision points so humans stay in control of interpretation.

On the technical side, the workflow ties together Azure-scale processing (via Premonition Insights) with external context from APIs such as Open-Meteo, and aims to keep analyses reproducible. If you build similar pipelines, this is a useful blueprint for structuring multi-step exploration so results can be reviewed, rerun, and audited instead of living only in ad hoc notebooks.

Other Machine Learning News

Several updates this week were about the plumbing that makes ML and analytics systems operational: secure connectivity, consistent runtimes, and observable data movement, which fits alongside last week's governance-and-ops thread by focusing on the connectivity and data-movement surfaces that agents and analytics pipelines ultimately depend on. If you are running hybrid data access or moving data between clouds and warehouses, these are the kinds of changes that reduce surprises in production.