Weekly ML Roundup: Fabric Agents, Secure AI, and GPU Workloads

by TechHub

Welcome to this week's ML roundup, covering confidential virtual machines on Azure Local, GPU-accelerated Fabric queries, and updates to Fabric's data and coding agents. We examine the upcoming query-generation rollout and November 6 Foundry integration migration deadline, then explore interactive visualizations, retrieval grounded in business data, and GPU memory profiling. Research reports highlight human-directed agents and explicit verification, while data-platform updates bring practical considerations for schemas, migrations, transactions, and governance, with availability and limitations distinguished throughout.

This Week's Overview

Azure infrastructure for local AI and GPU workloads

Microsoft's Azure updates cover confidential local execution, GPU-accelerated analytics, and driving dataset preparation. Availability differs across the announcements, with generally available capabilities alongside a public preview and an expected release.

Confidential local AI

Confidential virtual machines on Azure Local are generally available, using hardware-encrypted memory and attestation to protect data from host environments and infrastructure administrators. Foundry Local on Azure Local is expected to become generally available at Microsoft Ignite, so teams should distinguish the available infrastructure protection from the planned local AI release.

For developers handling sensitive data, the concrete security change is protection against access from the underlying host environment. Deployment planning should account separately for that protection and the availability of the runtime used to execute AI workloads.

GPU-accelerated warehouse queries

GPU-accelerated query acceleration for Microsoft Fabric Data Warehouse is in public preview and requires no changes to data or SQL queries. UNC Health reports queries running up to 5x faster, while WTW reports complex workloads running 3.4x faster at single concurrency.

Those results provide workload-specific evidence rather than a universal performance expectation. Teams can evaluate the preview against existing queries without first rewriting their SQL or reorganizing data.

Driving dataset preparation

NVIDIA Cosmos Data Factory's AV Dataset Generation blueprint becomes generally available on Azure. It supports curated, searchable, augmented, and evaluated driving datasets, covering preparation work before those datasets are used downstream.

Fabric data agents: better query generation and a migration deadline

Fabric data agents have a query-generation update approaching rollout and an integration retirement that requires action. Teams using them through Microsoft Foundry need to plan for the support deadline separately from the runtime improvements.

Advanced SQL and DAX generation

The generally available advanced natural-language-to-SQL and DAX generation update will begin reaching the standard runtime on November 2, with completion expected November 10. The rollout includes a move to GPT-5.6 Terra, while advanced SQL generation explores schemas, tests joins, and looks up actual filter values before producing queries.

Internal SQL-generation testing increased accuracy from 78.8% to 90.9% and reduced P95 latency, the 95th-percentile response time, from 46.8 to 32.9 seconds. Internal DAX testing increased accuracy from 89.15% to 91.86% and reduced P95 latency from 97.5 to 34.7 seconds.

These are internal evaluation results, so developers should validate generated queries against their own data and expected answers. Schema descriptions and topics remain available only for SQL sources on the preview runtime, an important distinction when comparing preview and standard-runtime behavior.

Foundry custom-tool retirement

Support for Fabric data agents connected as custom tools in Microsoft Foundry ends November 6, 2026. Affected Foundry agents must migrate to the preview Fabric IQ (OneLake Catalog) tool before that deadline.

The replacement supports multiple Fabric data agents per Foundry agent, compared with one in the earlier integration. OneLake Catalog lets users select accessible published agents by name without entering workspace or artifact IDs, while SDK and REST migrations connect Model Context Protocol (MCP) endpoints using the signed-in user's Microsoft Entra token.

The migration therefore changes both agent selection and the integration path. Teams should check their existing connections and authentication flow rather than treating the change as a tool-name replacement.

AI-generated visualizations that support interaction

Flint-interactive and Fabric's custom dashboard visuals address different parts of AI-assisted visualization. Flint defines how users interact with generated charts, while Fabric's preview generates a query and visual together inside a dashboard.

Flint-interactive connects chart actions to application state

Flint-interactive lets agents specify interactive chart behavior through semantic events and high-level update operators. Flint-MCP and MCP Apps render those charts within chat or canvas environments, and Flint Canvas operates as a state machine that manages user interactions and visual rendering state.

Applications can handle semantic events to update tables, other charts, widgets, or application state. Custom interactions combine triggers and handler functions with operators such as set style, set annotation, and set viewport.

For developers, this provides a way to connect an AI-generated chart to the surrounding application rather than treating it as a static image. The explicit event and state model gives applications defined points at which to respond to user actions.

Copilot generates custom Real-Time Dashboard tiles

Custom visuals with Copilot is in preview in Microsoft Fabric Real-Time Dashboards, generating a supporting Kusto Query Language (KQL) query and an HTML-based visualization as a new tile. Authors need editing access and data sources configured for custom visuals, and they can refine both outputs through follow-up prompts in the tile editor's Copilot pane.

Generated tiles still need validation against operational requirements. Authors should check aggregation, labels, and presentation alongside the query itself, rather than accepting a plausible-looking visual as evidence of correctness.

Agents in Fabric development and database investigation

Fabric's agent workflows now include live notebook execution and database investigation. Both give assistants access to operational context, but the notebook toolkit is a preview tool while the database coverage distinguishes demonstrated workflows from prototypes and roadmap plans.

Coding agents work with live notebook sessions

Fabric Notebook Toolkit (FNTK) is a preview command-line interface that lets coding agents discover notebooks and lakehouses, edit cells, execute code, and reuse an active notebook session. Structured JSON responses identify results, errors, and suggested next actions, while cell outputs, Spark job state, and driver, executor, and Livy logs provide debugging context.

Microsoft's data engineering team used GitHub Copilot with FNTK and connected a developer plugin to Azure DevOps through MCP. The practical distinction is that agents can inspect execution results and logs as they work, rather than stopping at generated notebook code.

Database agents support investigation, with autonomy still on the roadmap

The database agent observes SQL Server databases within Fabric Database Hub and surfaces performance signals and investigative context. Patrick LeBlanc demonstrates a blocking issue moving from detection in Database Hub to troubleshooting in SQL Server Management Studio, with GitHub Copilot assisting the deeper investigation.

Prototype skills target estate-wide analysis in VS Code. Outcome verification, memory, greater autonomy, human approval, and auditable controls are roadmap themes, not capabilities developers should assume are already available.

Retrieval grounded in application and business data

The retrieval examples connect embeddings, which represent data as vectors, with operational records and internal documents. They show distinct implementation paths in Azure Cosmos DB, Fabric SQL, and .NET.

Cosmos DB combines semantic search with live transactions

A travel-planner example stores operational application data and vector embeddings together in Azure Cosmos DB. Full-text search handles precise queries, vector search finds semantically similar results, and hybrid search balances semantic relevance with exact-match constraints.

Transactional workflows keep recommendations aligned with live inventory and booking data. That connection matters for applications where a relevant recommendation must also reflect what is currently available.

The Fabric SQL walkthrough covers database creation, migration of an existing SQL workload, and Git-based development and deployment across Fabric and VS Code. Automatic mirroring makes operational data available in OneLake without a separate replication pipeline.

Its AI example generates embeddings, runs vector similarity searches, and implements a T-SQL retrieval-augmented generation (RAG) pattern for product search. RAG grounds generated responses in retrieved data, making the example relevant to teams building search over existing business records.

.NET building blocks for ingestion and retrieval

Bruno Capuano and Luis Quintanilla cover Microsoft.Extensions.AI, Microsoft.Extensions.DataIngestion, and Microsoft.Extensions.VectorData. The session connects embeddings and vector search to context engineering and RAG grounded in internal documents and business data.

Tracing GPU memory pressure to application code

AKS GPU memory profiling is in preview, tracking GPU memory allocations rather than GPU compute utilization. Inspektor Gadget traces CUDA allocation calls at node level without application instrumentation or sidecar containers, then sends results to Pyroscope for display as Grafana flame graphs.

Developers can correlate those profiles with managed Prometheus metrics to investigate which code accounts for memory pressure. In the vLLM example, approximately 55.1 GiB of self allocation is attributed to the function that creates key-value (KV) cache tensors, illustrating the difference between seeing high memory use and identifying its source.

Human-directed agents in research

The mathematics and biological-monitoring reports show agents supporting research under human direction. Both retain explicit checks on their outputs, through proof verification in one case and experimental design and classifier evaluation in the other.

A dimension-independent proof of the KLS conjecture

Researchers including Microsoft Research's Xinzhi Zhang proved the KLS conjecture, establishing a universal bound independent of dimension. Human-directed AI agents explored more than 100 candidate approaches and assisted with derivations, manuscript preparation, and internal proof checks.

The October 1, 2026 paper version achieved iterated logarithmic dimension dependence, and the October 4 version removed dimension dependence. Repeated unguided runs with leading AI models using roughly the same total token budget did not solve the problem, distinguishing this result from autonomous discovery.

Independent proofs appeared, and a Lean 4 formalization was released. The authors also verified a streamlined argument in Lean, adding formal verification to the reported internal checks.

Biological monitoring combines synthetic data with experimental controls

Discovery agents interpreted over 120 software projects and reference documents through read-only access, compressing literature review, experimental design, permitting review, and preliminary classifier development from several weeks into a few days. Synthetic Asian citrus psyllid biometric data supported an initial edge classifier before permits allowed testing with live specimens.

The team selected a replicated Latin square design to balance sampling locations and periods in field experiments. A preliminary dataset assessed false positives against other insect species using simulated wingbeats and labeled laboratory recordings, keeping classifier evaluation separate from the initial synthetic-data development.

Fabric data engineering, analytics, and governance

Fabric's data-platform updates cover event schemas, warehouse migration, query expressiveness, and policy enforcement. Their practical boundaries matter as much as the additions: schema validation can drop events, object migration does not synchronize rows, and policies do not grant access.

Eventstreams adds schema-aware processing

Schema-aware Eventstreams is in preview for newly created Eventstreams, while existing ones continue operating unchanged. A single stream can handle registered, classified, and untyped events, with preview and processing configuration reflecting each event's shape.

Registered events use CloudEvents and the ce_type header to identify the event type, while field-based classification organizes events without creating registered schema contracts. Events from schema-associated sources that fail validation are dropped and reported in runtime logs, not routed as untyped, so developers need to monitor validation failures explicitly.

The broader Eventstreams update makes Schema Registry and Event Schema Sets generally available, with Avro schemas, version retention, change comparison, and discovery across workspaces. This separates the generally available schema-management capabilities from the preview processing behavior that uses them.

Warehouse migrations become incremental and add Teradata

Migration Assistant's preview supports moving database object definitions across multiple runs without requiring a new destination Warehouse. Dependencies are included by default, which can expand the selected scope, and naming conflicts require an explicit choice to keep, replace, or rename an object.

Supported sources include SQL Server databases, Azure Synapse Analytics dedicated SQL pools, and Teradata databases. The Teradata preview translates eligible metadata and SQL/BTEQ code into Fabric-compatible T-SQL, mapping macros to stored procedures and volatile tables to local temporary tables.

Object migration does not continuously synchronize changed rows, and the assistant does not move associated data. Data loading remains a separate step through options including a Fabric Data Factory copy job, staged files in Azure Data Lake Storage Gen2 with COPY INTO, or Fabric pipelines.

Graph queries gain more control over paths and stages

Fabric Graph's recap covers Graph Query Language (GQL) improvements introduced from January through August 2026, rather than a single new release this week. ANY SHORTEST selects one shortest path, path modes control repeated nodes or edges, and NEXT passes results between query stages.

EXISTS, CALL, and OPTIONAL CALL support contained graph queries, alongside filtered, paged, and nested aggregations and collection functions. Unbounded path patterns remove fixed maximum hop requirements, but they do not provide unlimited compute, so broader traversals still need resource-aware evaluation.

Warehouse functions distinguish exact and approximate distributions

New analytical functions are in preview for Fabric Data Warehouse and the SQL analytics endpoint. MEDIAN and QUANTILE provide exact calculations, while APPROX_MEDIAN and APPROX_QUANTILE estimate distribution metrics for large datasets.

PERCENTILE_CONT and PERCENTILE_DISC now support aggregate syntax alongside their existing window forms. Developers can return one result per dataset or group with aggregates, or preserve individual rows while adding analytical results with window functions.

T-SQL adds shorter grouping and ranking patterns

Fabric Data Warehouse now offers GROUP BY ALL, ORDER BY ALL, QUALIFY, and FROM ... SELECT in preview. GROUP BY ALL groups by every non-aggregate selected expression, while ORDER BY ALL sorts by selected columns in projection order.

QUALIFY filters rows after window functions are evaluated, supporting ranking and top-N patterns without another subquery or common table expression. FROM ... SELECT adds a source-first format that places selected columns after the data source.

Policy Sets add governance without granting permissions

Fabric Policies is in preview, supporting tenant- and capacity-scoped governance through Policy Sets, with only one active set per tenant or individual capacity. Advanced policies block actions unless every condition in at least one Allow rule matches, but they do not grant Fabric roles, workspace access, or item permissions.

Item creation policies cover Fabric item types, although Power BI item creation is not yet enforced. External data sharing controls are planned, so administrators should distinguish current enforcement coverage from future controls.

Other Machine Learning News

ABFS Vectored Read combines nearby byte ranges and queues requests asynchronously for compatible selective Parquet reads, requiring Apache Hadoop 3.3.5 or later with hadoop-azure and parquet-hadoop 1.12.3 or later. Spark workloads can enable it before startup with spark.hadoop.parquet.hadoop.vectored.io.enabled=true, but Azure Databricks does not currently support it, and evaluation should compare query duration, storage transactions, bytes read, and resource utilization because fewer requests can mean reading more intervening bytes.

Azure Databricks multi-table, multi-statement transactions require Unity Catalog managed Delta Lake or Apache Iceberg tables enabled for catalog commits, and external clients must retry the entire transaction against fresh data when optimistic concurrency checks reject a commit. Non-interactive transactions use row-level conflict detection, while interactive transactions use table-level detection except for unconditional appends, making transaction mode relevant to retry behavior.

Also relevant this week

Dates to know

Deadlines

Coming up