Weekly AI Roundup: .NET 11, Durable Agents and Foundry Models
.NET 11 is available with an SDK template for Model Context Protocol (MCP) servers, giving developers a shipped starting point for exposing application capabilities to agents while accompanying Blazor AI components remain experimental. The Durable Task extension for Microsoft Agent Framework reached general availability, allowing workflows to persist execution state, reuse recorded model results during replay, and wait for external approvals without keeping workers active. Azure Content Understanding added GPT-5.6, GPT-6, and GPT-6.1 model choices plus preview agentic extraction, while Microsoft Foundry expanded text and video options with distinct hosting, retention, and access conditions. Azure Speech made multilingual refinement and multichannel transcription generally available, although their compatibility limits mean developers cannot combine all features. Across these releases, security research and reliability evaluations reinforce the need to verify model provenance, constrain agent permissions, and validate actual state changes rather than trust an agent's apparent success.
This Week's Overview
- .NET 11 gives AI applications a new baseline
- Content Understanding separates model selection from extraction strategy
- Foundry model choices bring different hosting and access conditions
- Database MCP access needs explicit permissions
- AI-assisted Azure operations expand beyond queries
- Real-time speech adds coverage, channels, and agent handoffs
- Durable agent execution reaches general availability
- Foundry agent engineering focuses on state, identity, and release control
- Fabric agents depend on permissions and preserved business context
- Agent security reaches into model weights and worker boundaries
- AI-assisted vulnerability discovery still requires substantial validation work
- Local AI expands across Windows and disconnected deployments
- AI infrastructure targets preparation, storage, and traffic control
- Aspire examples make agent coordination observable
- Agent testing checks persistence and service behavior
- Reliability studies challenge apparent success
- Accountable agents leave authoritative actions to application code
- Shared agent skills need installation controls and ownership
- Copilot Studio updates connection setup and customization
- Other Artificial Intelligence News
- Also relevant this week
- Dates to know
.NET 11 gives AI applications a new baseline
.NET 11 is available as an STS release with two years of support, including an SDK template for Model Context Protocol (MCP) servers. The release sits alongside demonstrations of tool-connected applications and experimental interfaces, but developers need to distinguish shipped platform features from preview AI components.
Release availability and migration boundaries
The .NET 11 SDK is available to download, while .NET 8 and .NET 9 reached end of support on November 10, 2026. C# 15 adds union types and closed hierarchies, but its separate memory-safety work remains in preview. .NET MAUI now defaults to CoreCLR across platforms, with Android Native AOT still experimental.
For AI application teams, the SDK's MCP server template provides a starting point for exposing application capabilities to agents. Runtime Async and Blazor AI components remain preview or prerelease capabilities, so adopting the runtime does not make every accompanying AI feature production-ready. A companion overview explains how model abstractions, retrieval, tools, agents, local inference, and infrastructure fit together in the .NET AI stack.
MCP server implementation in C#
The MCP C# SDK session demonstrates scalable server implementation using stateless requests while retaining support for existing clients as the protocol evolves. It also covers collecting user input across multiple interactions, a useful distinction from tools that complete their work in a single request.
Tool-driven MAUI applications
A .NET MAUI demonstration connects an assistant to an existing conference app so users can find sessions, update an agenda, and navigate screens. The integration reuses application services and view models, keeping conversation and interface state aligned rather than creating a separate assistant-owned state model. It uses Microsoft.Extensions.AI and experimental .NET MAUI Labs tooling, with validation, confirmation before destructive actions, and handling of unsupported requests built into the discussion.
Experimental Blazor agent interfaces
The Blazor session pairs experimental AI components with the open AG-UI protocol to demonstrate streaming interactions, custom tool experiences, and human-in-the-loop workflows. Microsoft Agent Framework, Microsoft Foundry, and Aspire support examples that share application state and expose agent progress and tool activity to users.
A hardware companion backed by .NET services
The M5Stack/StackChan demonstration separates C++ hardware control from a .NET 11 backend, with Aspire orchestrating services and exposing telemetry. Microsoft Foundry supplies chat and vision, Microsoft Agent Framework supports agent behavior, and GitHub Copilot SDK handles tool-driven developer scenarios. The project includes sample code and setup guidance, while its WorkIQ demonstration uses synthetic or sanitized work context.
Content Understanding separates model selection from extraction strategy
Azure Content Understanding adds model choices and a preview workflow for documents that require more than a single extraction pass. These are separate decisions: teams can compare models under the same analyzer, then decide whether a workload warrants tool-driven extraction.
Model comparisons expose cost and quality trade-offs
Content Understanding now supports GPT-5.6, GPT-6, and GPT-6.1 across document, text, speech, and video workflows. In the document evaluations, GPT-6.1 Sol scored 95.74% on schema-only GA extraction, close to Astra's 95.95%, while Astra incurred about five times Sol's completion-token expense.
Examples can matter alongside model choice: five labeled examples increased Luna's extraction quality from 91.82% to 94.16%. Request-level modelDeployments mappings override resource defaults, allowing developers to compare models using the same analyzer and inputs rather than changing several variables at once.
Agentic extraction adds tools, tokens, and latency
Preview agentic mode can navigate document evidence, inspect visuals, calculate values, and validate structured output. Standard mode remains the default, faster, lower-cost option, while agentic mode typically takes longer and consumes more model tokens.
Configuration requires API version 2026-06-01-preview and workflow set to agentic under config. Content Understanding Studio exposes the same choice as Agentic Extraction when the corresponding preview API is selected.
Foundry model choices bring different hosting and access conditions
Microsoft Foundry's model announcements span text generation and video, with material differences in hosting, pricing, and eligibility. Developers need to evaluate those deployment conditions alongside model capabilities.
Claude availability includes restricted access and retention requirements
Claude Fable 5.2 and Fable 5.2 Fast are generally available, hosted on Anthropic with Global Standard and US DataZone deployment options. Fable 5.2 costs $8 per million input tokens and $40 per million output tokens. Fast doubles those prices and does not offer batch pricing, while Mythos 5.2 remains restricted to authorized customers whose eligibility is confirmed through Microsoft account teams.
Both models require safety-monitoring data retention, although eligible customers can use Fable 5.2 with zero retention until Enterprise Frontier Safeguards becomes available. Blocked cybersecurity queries receive responses from Opus 4.8, and blocked biology queries receive responses from Opus 5, so applications should account for those response paths rather than assuming every request uses the selected model.
Kimi combines a long context window with a choice of session owner
Kimi K3.1 Pro is an open-weight model with 2.8 trillion total parameters and a one-million-token context window. Fireworks supplies optimized inference infrastructure, while Foundry handles enterprise deployment and governance. Global Standard pricing is $2.05 per million uncached input tokens, $0.20 for cached input, and $10.35 for output.
The API choices include Chat Completions, Foundry Responses, and Fireworks Responses direct passthrough. That choice affects state ownership: Foundry Responses manages sessions in Foundry, while direct passthrough manages them in Fireworks.
US-hosted video models enter public preview
Five fal-powered video models are rolling out as a first-party Azure service in Foundry: Kling v3 Pro, LTX-2.5 Fast, LTX-2.5 Pro, MiniMax H3 Max, and MiniMax H3 Max Turbo. The initial portfolio is US-hosted, and Foundry is the first provider to offer a US-hosted Kling 3 endpoint.
LTX-2.5 Fast targets rapid iteration, while Pro prioritizes high-fidelity output, and both support synchronized audio. Developers estimating costs should resolve an inconsistency in the announcement before budgeting: its tables label rates per video second, but its introductory pricing text refers to USD per minute.
Database MCP access needs explicit permissions
Database tooling makes more operations available to agents, but network placement and convenient configuration do not establish authorization. The examples show different defaults, from read-only access to anonymous mutation capabilities that need review before deployment.
DocumentDB starts with read tools enabled
The open-source Azure DocumentDB MCP Toolkit is in public preview, exposing MongoDB-compatible database operations to agents. Read tools are enabled by default, while write and management tools require explicit environment flags. Vector search and BM25 full-text search run through the aggregate tool.
Administrator-defined connection profiles support Microsoft Entra ID, hostname allow lists, and SCRAM-SHA-256 for local or non-Entra environments. The toolkit can run through npx and supports Claude Code, Cursor, Codex, Gemini CLI, and GitHub Copilot, giving teams several client options without requiring write access from the outset.
Private PostgreSQL still needs application-level controls
The PostgreSQL architecture guidance separates private database connectivity from the identity, authority, and intent of the calling application. It compares public MCP, private MCP, and public-gateway designs, and clarifies that Azure App Service VNet integration provides outbound connectivity rather than private inbound access.
Gateway-backed deployments must block alternative public paths to the MCP backend or callers can bypass Azure API Management policies. Restricted workload permissions and database row-level security provide additional controls independent of MCP tool authorization.
SQL auto-configuration favors convenience over safe production defaults
Data API builder 2.0 introduced autoentities wildcard configuration, and its bundled container configuration exposes every table in every schema. The deployment examples use image version 2.0.12, with DAB_CONNSTRING overriding the primary configuration's connection string, including existing values or function references.
The default configuration enables development tooling and anonymous create, read, update, and delete operations. That makes it unsuitable for production without a permissions and exposure review.
AI-assisted Azure operations expand beyond queries
Azure management tooling now covers resource changes, deployment tracking, and cost investigation. The operational distinction is whether an agent can execute changes or only inspect data.
ARM tools add deployment execution and resource actions
The Azure Resource Manager MCP server supports resource and resource-group creation and updates through ARM PUT patterns, plus provider-defined POST actions through execute_resource_action. Deployment tools cover WhatIf previews, execution, status tracking, and cancellation, with Bicep MCP available for authoring and compilation. The get_resource tool retrieves current state for tracked, proxy, and extension resources by resource ID.
Cost tools add general and Azure Kubernetes Service (AKS) queries, retail price lookups, and negotiated price-sheet downloads. Experimental skills return SKILL.md workflows and require the request header x-mcp-toolset: Skills, keeping that capability explicitly opt-in.
FinOps toolkit adds read-only agents and data corrections
FinOps toolkit 15 adds an agent plugin with five agents and a read-only Azure MCP server scoped to Data Explorer. Local FinOps hubs require Docker with at least 16 GB of container memory, PowerShell 7, and FOCUS cost exports in Parquet format. Private deployments gain optional NAT Gateway support alongside VNet and public networking modes.
Teams upgrading should reingest historical Prices data to apply corrections to swapped commitment-discount eligibility columns. Azure Optimization Engine now uses a Data Collection Rule and Ingestion API following the Data Collection API's September 2026 deprecation.
Real-time speech adds coverage, channels, and agent handoffs
Azure Speech's generally available transcription updates serve different workloads and have different compatibility limits. Foundry's native voice-subagent preview adds a separate orchestration option, but it should not be confused with the production status of the transcription features.
Multilingual refinement expands locale coverage
Multilingual post-stream refinement is generally available, using an LLM Speech streaming model whose final results typically arrive in around 100 milliseconds. Coverage grows from 29 to 61 locales, and phrase lists support up to 200 phrases and 3,200 characters. Existing integrations receive the new engine without code changes, while multilingual post-stream refinement requires Speech SDK 1.50 or later.
The feature is available in Central India, East US, North Europe, Southeast Asia, West US, and West US 2. Monolingual post-stream refinement expands from 29 to 148 locales while retaining its parallel second-pass recognition architecture.
Multichannel recognition has a distinct feature set
Real-Time Multichannel Speech-to-Text is generally available, supporting up to two independent channels in one continuous recognition session at the same price as single-channel real-time transcription. It supports language identification per channel, and customers have run more than 3 million sessions since the August public preview.
Speech SDK 1.51 or later is required, with Go support and channel metadata on JavaScript diarization results requiring 1.52 or later. Phrase lists, multilingual models, post-stream refinement, and pronunciation assessment are unsupported, and heavy speech overlap can increase latency. Teams therefore cannot assume the multilingual refinement announcement applies to multichannel sessions.
Voice subagents separate conversation from backend work
Foundry's native voice subagents are in preview, with a frontline agent referencing a prompt or hosted backend agent by name in the same project. Linking currently requires the SDK rather than the portal, and the sample enables immediate acknowledgment only, not gap filling or progress narration.
Caller input does not automatically revise an in-flight backend task, and stopping audio playback does not establish task cancellation. Microsoft does not recommend the preview for production workloads.
Durable agent execution reaches general availability
The Durable Task extension for Microsoft Agent Framework is generally available, storing session and execution state outside the worker process. Completed agent calls with recorded results reuse those results during replay instead of calling the model again. Workflows can also wait for external approval events without keeping a worker active.
Applications still need to persist their lightweight session references and use the same Durable Task backend across requests. C# and Python applications can use Azure Functions or self-hosting, with execution visibility through the Durable Task Scheduler dashboard. A separate background-agent introduction covers autonomous execution and long-running work, while the extension announcement provides the concrete persistence and replay requirements.
- Durable Task extension for Microsoft Agent Framework is now Generally Available
- Your AI Agent Never Sleeps! 🔥 Background Agents Explained
Foundry agent engineering focuses on state, identity, and release control
Foundry examples examine what moves into the managed platform and what remains the application's responsibility. Hosting, tool deployment, and evaluation each retain explicit configuration and validation requirements.
Hosted agents reduce infrastructure ownership, not evaluation needs
A medical research-assistant migration moves from Azure Container Apps, PostgreSQL conversation storage, and dynamic sessions toward Foundry hosted agents. Hosted agents provide VM-isolated session sandboxes and manage compute, persistence, scaling, endpoints, and agent identity.
The team preserved telemetry in both Azure Monitor Application Insights and Langfuse, and stable backend contracts allowed reuse of existing evaluations against expert-supplied golden answers. Networking guidance recommends a /24 subnet for production, with /27 documented as the supported minimum for smaller deployments. A companion C# session explores moving .NET agents from localhost into the managed environment.
- What it takes to bring an AI research assistant to thousands of medical experts
- Beyond Localhost: .NET Agents in Foundry
A managed harness experiment exposes grounding gaps
An experiment replaces four hosted Microsoft Agent Framework wrappers with one Foundry-managed Copilot harness agent, while a React workbench and Node.js backend retain authentication, streaming, conversation ownership, and authorized downloads. Fabric access uses the signed-in user's context, whereas Foundry IQ uses the project's managed identity.
Early tests returned correct sales figures without retrieving the requested explanatory document, and clarified instructions produced one successful retrieval from both sources. The harness remains in public preview, and the demonstration does not establish production readiness, feature parity, or better quality, latency, or cost.
Toolbox versions require deliberate promotion
A Foundry Toolbox exposes a versioned collection of tools through one MCP-compatible endpoint within a Foundry Project. Versions are immutable configuration snapshots, subsequent versions require explicit promotion, and rollback means promoting a retained earlier version.
The project, project-level role-based access control, and required connections must exist before deployment through the project endpoint. Guidance requires Azure Developer CLI 1.37.0 or later and the unified microsoft.foundry extension. The sample promotes immediately through a postdeploy hook, but production releases should validate candidates first.
Tool selection and delegated access need separate controls
Foundry tool search retrieves relevant tools instead of loading every definition into every model request. User delegation supports On-Behalf-Of access so downstream services can authorize requests using the user's identity, while evaluation datasets can come from sampled traces or be synthesized from instructions, prompts, and reference documents.
Insights, rubric evaluators, custom evaluators, and managed memory are described as preview capabilities. AI Gateway enforces token-rate limits and total token quotas, not native dollar-budget enforcement in Agent Service.
Fabric agents depend on permissions and preserved business context
Fabric-related guidance connects natural-language investigation with the data definitions and values that make answers meaningful. The developments include a read-only Copilot preview, a governance retirement, and an evaluation showing how language differences can break filters.
Database Hub Copilot enters public preview
Copilot in Database Hub answers questions using resource inventory, detected issues, and performance telemetry. It respects existing Azure and Fabric permissions and does not change databases, making this an investigation interface rather than a database mutation tool.
A Fabric administrator must enable the required tenant settings, and access requires Fabric F2 or higher or Power BI Premium P1 or higher. Token usage consumes capacity units, and OpenAI-operated models are not currently available in government or sovereign clouds.
Purview retirement makes context inventory a migration task
Microsoft announced the retirement of Purview Data Governance on November 12, 2026, with support ending on November 12, 2027. The article identifies OneLake Catalog as the center of Microsoft's data governance, but argues that migration should begin by inventorying definitions, ownership, lineage, rules, classifications, and relationships rather than choosing a replacement catalog first.
Fabric IQ's Ontology feature binds business concepts and relationships to data for AI consumption. The authors recommend governing reusable business terms outside individual platforms and serving that context through Fabric ontologies, preserving meaning as the underlying technology changes.
Multilingual evaluation finds value-normalization failures
An exploratory test of a synthetic 288-row Fabric warehouse table graded 42 responses and traced all 10 incorrect answers to Traditional-script filter values absent from the stored Simplified Chinese values. A Foundry normalization route produced seven correct answers and three incomplete runs, with no incorrect completed answers.
The result is narrow, and its implementation path has a lifecycle constraint: Microsoft Learn lists November 6, 2026 as the Fabric data agent tool's deprecation date and directs users to Fabric IQ preview. Fabric IQ was not tested, so the article recommends repeating independent SQL-based evaluations after migration.
Agent security reaches into model weights and worker boundaries
Security research examines backdoors introduced before deployment, while implementation guidance limits what an agent can access during execution. Behavioral tests, model provenance, and runtime isolation address different parts of that risk.
Post-training poisoning can survive restored coding performance
Researchers inserted 90 poisoned examples into roughly 9,000 terminal-agent conversations across 10 mock harnesses. Fine-tuned GPT-4.1 mini produced unsafe actions in 78.0% and 77.8% of trials on seen harnesses, but transfer to unseen harnesses fell to 11.0% and 0%.
The strongest model-poisoning configuration achieved 84.1%-94.8% backdoor activation on unseen repositories using Qwen3.5-4B and Qwen3.5-9B. Reinforcement learning recovered much of the models' benign coding performance without materially removing the backdoor success. These results require a backdoored checkpoint and do not demonstrate weaknesses in default Claude Code or Codex CLI configurations.
Weight-level backdoors need weight-level inspection
A separate study shows that edits to fewer than ten neurons can create a sleeper agent that passes evaluations but emits an attacker's command when triggered. Implanting the backdoor requires only enough RAM to run the model, although Microsoft has not observed threat actors weaponizing this technique.
Microsoft recommends auditing weights before deployment and after every update, alongside behavioral backdoor scanning. Defender for Cloud model scanning covers serialization and code-level risks, but does not replace weight-level audits.
Isolated workers keep credentials outside agent execution
Patrick Nikoletich demonstrates a Dockerized worker pool with isolated containers and credential-safe inference gateways that avoid exposing inference credentials to workers. Agent Host Protocol tunnels CLI sessions into those containers, and the repository includes a harness builder and secure Aspire agent service.
Copilot Harness 2.0's Rust rewrite and direct Azure tunneling are plans, not demonstrated capabilities. A related discussion with Niels Provos explains why unrestricted access to workplace tools and potentially malicious third-party skills makes these execution boundaries important.
- Inside the Microsoft Copilot Harness: Customizing AI Agents
- The hidden security risk of running AI agents without boundaries
AI-assisted vulnerability discovery still requires substantial validation work
MDASH-related research reports findings in Linux and Rust, but the work extends well beyond scanning. The reports distinguish validated issues, upstream fixes, and practical exploitability rather than treating every finding as an equivalent security outcome.
Linux reports quantify the downstream workload
MDASH scanning produced 26,502 raw findings, narrowed to 621 proof-of-concept-backed cases and 59 separate upstream reports. Those reports received 33 CVEs, including 24 rated High or above under CVSS v3.1. Fixes for 38 reports had merged upstream, with a median of 41 days from first public patch to mainline release.
CVE-2026-64563 enabled stable local privilege escalation on Linux and affected Android with SELinux enforcing, without special kernel configurations or user namespaces. Reporting the 59 validated issues required more than four person-months after approximately two weeks of scanning, showing how much expert work remains after automated discovery.
Rust findings examine soundness assumptions
FORGE Lab investigated Rust standard-library soundness bugs involving panic safety, user-provided implementations, and operating-system state. Some findings came from Hyenas, FORGE's frontier research version of the AI-powered scanning system codename MDASH. Issues affecting strip_prefix_of and split_once remained open on November 9, 2026.
Rust 1.100, scheduled for November 12, requires Allocator implementors never to unwind from trait methods or drop. FORGE assessed exploitation of the signal-stack cleanup bug as unlikely under tested conditions, but did not extend that conclusion to its other findings.
Local AI expands across Windows and disconnected deployments
Local inference updates cover model rollout, runtime APIs, and deployment environments. Availability differs across these layers, with planned Windows releases, experimental APIs, and a separate Azure Local announcement.
Aion Instruct requires application transition checks
Aion Instruct has 2.3 billion parameters and a 32K-token input context limit, with Windows Insider Preview rollout planned for November 20, 2026 and retail rollout planned for January 2027. GPU support is expanding to supported hardware, including NVIDIA RTX-series GPUs and AMD RDNA4 devices.
Applications using custom prompts, custom LoRAs, or custom output parsing require attention during the transition. The model also adds invisible watermarking to meet text-provenance requirements in specific regions, making output behavior part of the compatibility review.
Windows ML previews task-specific runtime APIs
Windows ML adds experimental llama.cpp support for local GGUF models, alongside a preview Runtime API whose first release includes Text Generation and Speech Recognition. Existing ONNX Runtime APIs remain fully supported, and an OpenAI-compatible endpoint supports local prototyping.
The Runtime APIs support zero-copy Windows-native inputs, deterministic multi-model pipelines, and ahead-of-time loading and compilation. Examples cover Triton kernel compilation with PyTorch, ONNX export, and preparation and benchmarking with Windows ML CLI. A separate Foundry Local session approaches small-model selection from C#, weighing capability, hardware speed, and cost without requiring a cloud connection.
Sovereign deployment includes disconnected datacenters
Thomas Maurer's session describes Foundry Local on Azure Local supporting connected or fully disconnected datacenter deployments with Microsoft, partner, open-weight, and open-source models. The article states that Microsoft announced general availability for Microsoft Ignite 2026, and places the deployment option alongside region selection, the EU Data Boundary, confidential computing, and Microsoft Purview controls.
AI infrastructure targets preparation, storage, and traffic control
Azure infrastructure updates address work around model inference as well as inference itself. The practical considerations include sandbox density, shared datasets, and failover behavior, with preview and benchmark limitations that matter for deployment decisions.
HDv2 opens preview registration for CPU-heavy AI workloads
Azure HDv2 preview registration is open for AI data preparation, compute-intensive high-performance computing, and dense agentic sandboxes. The machines use AMD EPYC 9006 processors with more than 500 physical Zen 6 cores, 3.8 TB of DDR5-8800 memory, and 400 Gb Azure Boost Networking.
Microsoft reports approximately twice as many parallel sandboxes and up to 1.7 times more agentic task throughput than the next-best in-market alternative. Those comparisons are preliminary, and SPEC CPU 2026 Integer Rate is a proxy for agentic throughput rather than a direct application benchmark. Related Azure and AMD guidance describes Foundry Managed Compute with dedicated MI300X capacity through private endpoints, Microsoft-managed serving software, and GPU-hour billing.
- Accelerating AI Data Processing, HPC, and Agentic Workloads with Azure HDv2 Virtual Machines
- Visit Azure and AMD at Microsoft Ignite and SC26
NetApp Files connects training datasets and inference storage
Azure NetApp Files supports checkpoint writes, high-throughput NFS reads, quality-of-service controls, and writable short-term clones for parallel experiments. Trident provisions persistent volumes for containerized inference on AKS, while the object REST API lets Microsoft AI and data services access NFS and SMB datasets without copying them into a separate object store. The reference deployments report qualitative improvements, not numerical latency, throughput, or training-time comparisons.
- Azure NetApp Files: The Data Backbone for NVIDIA AI Model Training, Inference, and Microsoft Foundry
Application Gateway previews weighted routing and failover
Azure Application Gateway's public preview adds routing rules with up to five AND-combined conditions, RE2 expressions, and backend-pool selection by weight or priority. Priority failover uses Minimum Healthy Percent and includes scenarios that move AI traffic from reserved PTU endpoints to pay-as-you-go endpoints.
API version 2026-03-01 or later is required. Preview weighted distribution ignores pool health, and the previews have no service-level agreement, so teams should test failure behavior rather than infer it from routing weights.
Aspire examples make agent coordination observable
Aspire examples address both running distributed agents and using agents to maintain distributed applications. In each case, explicit dependencies and deterministic checks provide evidence beyond the agent's final response.
Multi-repository rollout uses measured repository health
MARS combines Aspire, Microsoft Agent Framework, and GitHub Copilot SDK in a check, plan, fix, validate, and learn loop. By May 2026, more than 50 microservice repositories had adopted Aspire, up from approximately 15 in the previous report.
Repository health comes from deterministic C# metrics, with cloud checks grounded in Azure DevOps pull-request and build snapshots. The team estimates that agentic onboarding removes 4-8 hours of manual setup per repository, and every pull request against the rollout system runs an end-to-end suite of more than 700 tests.
Distributed specialists share one application model
AlpineAI brings .NET and Python specialist agents, a Go telemetry generator, a React frontend, and Foundry resources into one Aspire AppHost. Its advisor exposes four A2A specialists and a Foundry prompt agent as five tools, while chat uses a Foundry hosted agent and voice uses a .NET WebSocket service connected to Azure AI Voice Live.
Aspire models Azure Container Apps placement, Cosmos DB conversation storage, dependencies, and health checks. The dashboard shows OpenTelemetry traces across advisor processing, specialist calls, model calls, and downstream HTTP requests.
Agent testing checks persistence and service behavior
Local test environments can validate more than whether a model returns plausible text. The examples inspect stored state, isolate dependencies, and exercise real service graphs while keeping cloud-validation limits explicit.
Cosmos DB tests connect answers to persisted state
The generally available Cosmos DB vNext emulator runs through a Linux container on Windows, macOS, and Linux hosts across x64 and ARM64. Startup scripts mounted at /init run on first startup when ENABLE_INIT_DATA=true, not before every test, and OpenTelemetry Protocol export plus shell --otel can correlate client and gateway spans. Authentication is not enforced, so emulator tests do not replace validation against Azure.
The FleetCharge example uses separate GitHub Actions component and live-model jobs, each with a fresh emulator, and local Phi-3.5 inference through Ollama. Its live tests make 20 real inference calls and verify recommendation creation, replacement, and deletion through SDK checks. No Azure credentials are required, but repository rules or branch protection must require both jobs for an enforced merge gate.
- AI Agents: Meet the Azure Cosmos DB vNext Emulator
- From Local Development to CI Workflows with the Azure Cosmos DB Emulator
Hermetic Aspire tests exercise the real service graph
Azure Chaos Studio runs roughly 90 hermetic tests on every pull request in Azure Pipelines, in parallel without shared-environment interference. Emulators and WireMock replace dependencies while tests exercise zone outages, identity outages, DNS failures, and geo-replication failures across the actual service graph. The Aspire fixture waits for real health checks before testing and tears down the application afterward.
Reliability studies challenge apparent success
Two evaluations show why an agent's confidence or choice to stop is not enough to establish correctness. One measures failures that look complete, while the other isolates how dependency-version labels change review behavior.
ThinkingBox measures convincing failures
The expanded ThinkingBox study evaluates 18 models across 507 benchmark tasks with 20 attempts per task. Among 79,853 failed trials, 67.24% appeared complete based on final responses, state-changing tool calls, and error-free final tool responses.
After 50 reinforcement-learning steps, Qwen3.8-27B improved pass@1 from 51.70% to 60.78%, and tasks succeeding on all 20 attempts increased from 38 to 107. Estimated cost per all-20 task fell from $24.36 to $8.00, but excludes training, task-selection probes, simulated users, judges, and infrastructure.
SemVer labels influence dependency scrutiny
Researchers used Vally and GitHub Copilot CLI to test identical dependency code under different semantic-version labels. Claude Sonnet 5.5 checked 2 of 10 patch upgrades but all 10 major upgrades in the first experiment.
A broader review experiment covered 540 runs across six models, six scenarios, and patch, minor, and major labels. Claude Haiku 5.5 merged updates unchecked in 37% of patch runs versus 7% of major runs, and all four breaking-update merges that left main broken carried patch or minor labels. The findings support validating the actual change rather than allowing its version label to determine the testing requirement.
Accountable agents leave authoritative actions to application code
Architecture examples separate proposals from the services authorized to change state. That division gives applications a place to enforce policy, validate outcomes, and handle uncertain execution.
Scheduling agents propose placements, not calendar truth
The agentic scheduler lets agents propose placements while application code owns calendar state and validates demand, duration, spacing, conflicts, and capacity. Its Node.js backend runs on Azure Container Apps, using PostgreSQL queues and checkpoints plus an Azure Files workspace.
The gateway persists commands before returning 202 Accepted, and queued replies reach browsers through Server-Sent Events. Whole-grid evaluation targets failing programs for up to three retries, with model requests and tool calls traced through OpenTelemetry and Azure Application Insights.
Trusted services enforce policy and reconcile uncertain outcomes
An illustrative Azure architecture separates experience and reasoning, trusted action, and operational intelligence into three logical planes. Models propose actions, while trusted services check authorization, policy, and authoritative state before changing records. Unknown partner outcomes require reconciliation before retries because duplicate detection alone cannot resolve an uncertain booking.
The design keeps Fabric analytics outside booking, payment, and command execution, and proposes pilots moving from shadow mode to assisted booking and bounded autonomy. A related Logic Apps and Functions example applies the same separation to alerts: workflows collect context, Functions process and validate technical data, and generative AI recommends next steps subject to validation or human review.
- The Model Isn’t the Boss: An Azure Architecture for Accountable AI Agents
- From Automation to Intelligence: Logic Apps + Functions + GenAI
Shared agent skills need installation controls and ownership
Skill distribution and internal catalogs address complementary problems: approving what enters a workspace and making approved tools discoverable. Neither should be confused with proving that the installed behavior is correct.
Git-Ape makes source and destination approval explicit
Git-Ape skill installation requires Node.js, an authenticated GitHub CLI, and a client supporting workspace skills. Search uses the registry bundled with the installed copy, not an automatic query of the latest GitHub registry, and users approve the skill, pinned source commit, and destination before files are written.
Proposed pack installation verifies member content, refuses overwrites, and leaves skills independently usable. Installation does not run community scripts, install their dependencies, or validate generated deployment designs.
Slack's catalog turns personal configuration into shared infrastructure
Slack's internal AI tool catalog grew from Frances Coronel's personal setup after other engineers began forking it. The session demonstrates using GitHub Projects and Issues as a single source of truth for MCP servers, skills, and GitHub Copilot agents, with an adoption dashboard and six months of usage metrics.
Copilot Studio updates connection setup and customization
Copilot Studio guidance covers both how users establish third-party connections and how builders extend agent behavior. The connection changes have a stated future availability target rather than a current general-availability release.
In-conversation connections require custom-interface changes
Microsoft expects the improved connection experience to reach general availability by December 31, 2026. Users will configure third-party connections within conversations, with eligible valid connections potentially reused across agents accessed through Microsoft Copilot, Microsoft Teams, and Web Chat.
Custom web interfaces using the Standard harness must open the new connection window and resume the conversation after setup succeeds. Existing organizational controls, authentication requirements, and user permissions continue to govern access.
Hooks provide a customization entry point
A separate introductory video explains hooks as a way to extend Copilot Studio agents and customize behavior. It covers when to use them, practical applications, and agent-building guidance for developers and beginners.
Other Artificial Intelligence News
Visual Studio 18.12 makes Bring Your Own Model generally available, allowing hosted or local models in Agent Mode without GitHub sign-in, with separate provider configurations and support for OpenAI-compatible Completions and Responses plus Anthropic-compatible Messages. Connections use Visual Studio's resolved proxy configuration, and administrators can control availability through organizational policy.
The NVIDIA BioNeMo Agent Toolkit brings natural-language protein-binder design into Microsoft Discovery, with RFD4-Proteina generating structures and amino acid sequences in one model pass while a CPU-only Python 3.11 tool container holds neither model weights nor CUDA. Inference can use NVIDIA's hosted API, Foundry, or AKS, and molecular data stays within the customer's Azure tenant when both the container and model run on AKS inside that tenant.
Microsoft Web IQ demonstrates grounding in current web information through passage-level retrieval and citations, with scopes controlling source selection. The session covers Auto search, Places, fast versus advanced retrieval, and inspection of the evidence behind an agent's US energy-market analysis.
Azure's .NET demonstration covers isolated code execution in a Container Apps sandbox, model and tool tracking through AI Gateway, operational help from SRE Agent, and Copilot canvases for developer workflows.
The SQL roundup includes DiskANN vector search with data modifications and asynchronous index maintenance, plus generally available GitHub Copilot Agent Mode in SSMS 22.10.2. The SQL AI Developer resources map Exam DP-800 to SQL Server, Azure SQL, and Fabric SQL development, with learning material on embeddings, hybrid search, retrieval-augmented generation, and SQL MCP Server.
- What’s new across Microsoft SQL in 2026 so far: the PASS Summit West roundup
- Resources for the SQL AI Developer
Fabric's October community spotlight collects 46 posts, videos, and tutorials, including guidance on securing data agents and governing semantic models for AI. The collection also covers Fabric IQ, deployment plans, CI/CD, and a free DP-700 study course.
Adrian Twarog's Foundry walkthrough moves from account setup and model exploration to agent publication, including a scoped MCP connection and an Azure SDK call. Companion clips emphasize spending visibility and a single-prompt playground check, which is a focused pre-publication check rather than a comprehensive evaluation.
- Microsoft Foundry with Adrian Twarog: Build Your First Agent
- Microsoft Foundry: Keep Your AI Spend Under Control
- Microsoft Foundry: Test Your Agent Before You Ship
- Path to Production for AI Agents
Yunika Liu's Coco project connects a pixel-art desktop dog and LED display through a custom MCP server built with GitHub Copilot. The session covers WebSocket synchronization, animation triggers, music composition, and lessons from building a first MCP server in one week.
GitHub Universe discussions examine both what AI-assisted development may support and what developers should choose to build. Sue Black and Kent C. Dodds discuss fitness for purpose, maintenance, and wider effects, while a panel representing GitHub Copilot, Anthropic, Google Cloud, and OpenAI considers models, agents, security, architecture, and developer skills over the next two years.
- What's worth building when AI can build anything?
- The view from the labs: What's next for AI-assisted development
Also relevant this week
- Aspire 13.4 is here: Aspire 13.4 makes TypeScript apphost generally available, using apphost.mts as the entry point and running startup validation before the application. Full story in the .NET roundup.
- Plan Your Java Upgrade Ahead of Microsoft’s 2027 Support Milestones: Azure App Service and Azure Functions support for Java 8, 11, and 17 ends September 1, 2027. Full story in the Azure roundup.
- Introducing Azure Container Apps as a target workload in Azure Migrate: Azure Migrate adds Azure Container Apps as a target workload for assessing discovered web applications. Full story in the Azure roundup.
- Access GitHub Copilot in JetBrains – GitHub Copilot Is Now a Native Agent in JetBrains AI Assistant: The original integration makes GitHub Copilot a default option in JetBrains AI Assistant’s agent picker through ACP. Full story in the GitHub Copilot roundup.
- Claude Fable 5.2 in GitHub Copilot: Claude Fable 5.2 is available in GitHub Copilot Pro+, Max, Business, and Enterprise, with a gradual rollout. Full story in the GitHub Copilot roundup.
- Copilot in Visual Studio Just Got a Major Upgrade: The SDK-powered Copilot Chat experience is enabled by default in Visual Studio 2026 version 18.12. Full story in the GitHub Copilot roundup.
- Gemini 4 Carbon in GitHub Copilot: Gemini 4 Carbon is generally available in GitHub Copilot for Pro+, Max, Business, and Enterprise users. Full story in the GitHub Copilot roundup.
- Gemini 4 Hydrogen in GitHub Copilot: Gemini 4 Hydrogen is now generally available in GitHub Copilot, with improved instruction following over Gemini 3.8 Flash. Full story in the GitHub Copilot roundup.
- GitHub Copilot for Xcode: Unlocking Agentic Power for Apple Developers: GitHub Copilot brings Custom Agents, Subagents, and Plan Mode to Xcode. Full story in the GitHub Copilot roundup.
- GitHub Copilot is now a default agent in JetBrains Air: GitHub Copilot is available in JetBrains Air’s default agent picker through Agent Client Protocol, without a separate Copilot plugin installation. Full story in the GitHub Copilot roundup.
- GitHub Enterprise Server support in the GitHub Copilot app: GitHub Enterprise Server support in the GitHub Copilot app is generally available after an opt-in preview. Full story in the GitHub Copilot roundup.
- Important updates: GitHub Copilot support ending for JetBrains 2025.1.x: GitHub Copilot will discontinue support for JetBrains IDEs 2025.1.x, including all patch releases, in the next few months. Full story in the GitHub Copilot roundup.
- Important updates: GitHub Copilot support ending for JetBrains 2025.2.x: GitHub Copilot will end support for JetBrains IDEs version 2025.2.x, platform build 252, including patch releases, in the coming months. Full story in the GitHub Copilot roundup.
- Upcoming deprecation of selected GitHub Copilot models in December: GitHub Copilot will deprecate six models across all experiences on December 14, 2026. Full story in the GitHub Copilot roundup.
- Pylance + Pyrefly: bring your preferred type checker to Visual Studio Code: Microsoft and Meta designed and implemented Type Server Protocol, an open standard extending Language Server Protocol for interchangeable type checker backends. Full story in the DevOps roundup.
- Modern web development with ASP.NET Core & Blazor in .NET 11: The session covers ASP.NET Core and Blazor in .NET 11 through demonstrations. Full story in the .NET roundup.
- Deploying an AI Application on Azure Container Apps: The deployment architecture connects container images in Azure Container Registry to Azure Container Apps and Azure OpenAI. Full story in the Azure roundup.
- Agentic CI/CD with Azure Cosmos DB: Isolated Environments for a Swarm of AI Agents: The sample assigns each coding task a separate Git worktree, owned Azure Cosmos DB emulator, versioned fixtures, and task-specific artifacts. Full story in the DevOps roundup.
- Modernize MongoDB applications with Azure DocumentDB: From assessment to what’s next: Azure DocumentDB is a fully managed MongoDB-compatible service built on the PostgreSQL-based open-source DocumentDB engine governed by the Linux Foundation. Full story in the Azure roundup.
- Migrate NuGet-based Logic Apps Standard apps and custom built-in connectors to .NET 10: Migration preserves connector registration code and workflow definitions; projects must target net10.0 and run on a .NET 10 host. Full story in the Azure roundup.
- Introducing Azure Monitor dashboards with Grafana MCP: Telemetry for AI Agents, Signed In as You: The global MCP endpoint requires no Azure Managed Grafana resource, MCP server deployment or Grafana data source configuration. Full story in the Azure roundup.
- Azure VMs deliver inference performance in line with industry benchmarks: ND GB200 v6 testing covered 264 matched comparisons against bare metal, using four GPUs per node. Full story in the Azure roundup.
- GitHub availability report: October 2026: As of November 9, 140 of 151 production services and 32 of 50 MySQL primaries were running in Azure Central US. Full story in the DevOps roundup.
- Manage AI Scan for repositories using the REST API: The public preview adds repository_scan to organization and repository AI Scan REST API endpoints. Full story in the Security roundup.
- 3D Modeling with GitHub Copilot: From Brief to Shape: The projects use Python for measured geometry and Blender for organic shapes, with GitHub Copilot editing generators, scenes, model files, and documentation. Full story in the GitHub Copilot roundup.
- Build once, run on any agent: A practical guide to agent skills: An agent skill is a versioned directory centered on SKILL.md; its host must supply tools, credentials, and permissions. Full story in the GitHub Copilot roundup.
- Four days, zero lines: the AI force multiplier with GitHub Copilot: Jeff Doolittle used GitHub Copilot to build an automated governance system in four days without manually writing a single line of code. Full story in the GitHub Copilot roundup.
- GitHub App support for Enterprise Copilot seat management: Enterprise GitHub Apps can assign GitHub Copilot seats to users and enterprise teams through enterprise-level seat management APIs. Full story in the GitHub Copilot roundup.
- GitHub Copilot modernization: The session demonstrates a real .NET migration end to end using agentic modernization. Full story in the GitHub Copilot roundup.
- Python support has arrived in the guided Copilot experience for VS Code: Python projects now have full support in the guided Copilot experience for building Azure apps in VS Code. Full story in the GitHub Copilot roundup.
- Smarter Spring Development in Eclipse with GitHub Copilot: The setup requires Spring Tools for Eclipse 2.0 or later, the GitHub Copilot plugin, and an open Spring Boot project. Full story in the GitHub Copilot roundup.
- The State of AI in Visual Studio: The video explores a new Visual Studio agent harness powered by the Copilot SDK. Full story in the GitHub Copilot roundup.
- Visual Studio Code and GitHub Copilot - What's new in 1.144 + 145 - GitHub Universe 26 edition: The video demonstrates selected GitHub Copilot features in Visual Studio Code 1.144 and 1.145. Full story in the GitHub Copilot roundup.
- Visual Studio Code and GitHub Copilot - What's new in 1.146: The video demonstrates selected GitHub Copilot features in Visual Studio Code 1.146 rather than covering every release feature. Full story in the GitHub Copilot roundup.
- What is new in GitHub: October releases explained: The episode reviews October updates across the GitHub ecosystem. Full story in the GitHub Copilot roundup.
- I made my Octolamp think with GitHub Copilot CLI hooks: GitHub Copilot CLI lifecycle hooks update an Octolamp through the WLED API. Full story in the GitHub Copilot roundup.
- From Silicon to Service: Bringing AMD Helios™ to Azure: Each Azure ND MI455X v7 compute rack contains 72 MI455X GPUs, 18 AMD EPYC Venice CPUs, and 31 TB of HBM4 memory. Full story in the Azure roundup.
- From Purview Data Governance to the Fabric OneLake Catalog: The OneLake catalog organizes Fabric data discovery and governance into Explore, Govern, and Secure experiences. Full story in the Machine Learning roundup.
- Advancing the Quality-Efficiency Frontier for Coding AI on Windows PCs: MAI-Code-1.1-Flash is a 137.9-billion-parameter mixture-of-experts model with a released 53.0 GiB mixed-precision checkpoint. Full story in the Machine Learning roundup.
- What's New in Visual Studio Debugging and Diagnostics: The video demonstrates Visual Studio debugging and profiling experiences through practical demos. Full story in the .NET roundup.
- Python in Visual Studio Code – November 2025 Release: The November 2025 release of the Python extension for Visual Studio Code is available. Full story in the DevOps roundup.
- Why WinForms must not change - yet reinvented itself with .NET 11!: WinForms modernization in .NET 11 prioritizes preserving behavioral and designer compatibility for existing business applications. Full story in the .NET roundup.
- Challenges of building machine learning compilers across varying hardware architecture: Strong ML compiler performance requires progressive lowering across multiple intermediate representations, hardware-aware tiling, data movement optimization, and execution scheduling. Full story in the Machine Learning roundup.
- Building great Windows apps with WinUI and agents: The video covers new dotnet new templates for creating WinUI projects. Full story in the .NET roundup.
- How Aspire uses .NET to build an agent-first toolchain for any programming language: Aspire uses .NET as its orchestration, developer experience, and runtime foundation without restricting application languages to .NET. Full story in the .NET roundup.
- 3D printing a custom microphone stand with GitHub Copilot: The team used GitHub Copilot to modify a 3D model of a DJI microphone stand instead of manually editing it in CAD software. Full story in the GitHub Copilot roundup.
- 3D Printing with GitHub Copilot: Building the Workflow: Both repositories were created and maintained entirely through GitHub Copilot App sessions, including tools, generators, tests, documentation, and model deliverables. Full story in the GitHub Copilot roundup.
- C++ language server now auto-configures MSBuild builds: The Microsoft C++ Language Server now bundles an MSBuild compile commands extractor and invokes it automatically on project open. Full story in the GitHub Copilot roundup.
- Protect your PC from AI coding agents while modernising legacy .NET apps: The session discusses isolation options for AI coding agents using GitHub Copilot CLI modernisation plugins on legacy .NET applications. Full story in the Security roundup.
- Turning the GitHub Universe badge into a retro gaming console: The GitHub Universe badge uses graphics as capacitive touch buttons for handheld gaming controls. Full story in the GitHub Copilot roundup.
- VS Code: Release Recap: The session covers Visual Studio Code and GitHub Copilot updates from the past month. Full story in the GitHub Copilot roundup.
- Test Smarter: AI-Assisted Testing for Modern .NET: Session BRK038 demonstrates AI-assisted testing with MSTest, Microsoft Testing Platform, and modern .NET tooling. Full story in the .NET roundup.
- Visual Studio Code 1.147 (Insiders): The release notes cover Visual Studio Code 1.147 Insiders and were last updated on November 10, 2026. Full story in the DevOps roundup.
Dates to know
Deadlines
- 14 December 2026: GitHub Copilot: Claude Fable 5, Claude Haiku 4.5, Gemini 3.8 Flash, Gemini 4 Argon, Grok 4.6, and Kimi K3 are deprecated across Copilot experiences; individual annual-plan subscribers retain Claude Haiku 4.5. (Upcoming deprecation of selected GitHub Copilot models in December)
- 1 January 2027: GitHub Copilot: The temporary enterprise zero data retention exemption ends after 2026; continued use of Claude Fable 5, 5.1, and 5.2 requires EFS. (Claude Fable 5.2 in GitHub Copilot)
- 1 September 2027: End of support — Azure App Service; Azure Functions: Support for Java 8, 11, and 17 ends; existing applications continue running without related runtime updates or support. (Plan Your Java Upgrade Ahead of Microsoft’s 2027 Support Milestones)
- September 2027: Microsoft Build of OpenJDK: Earliest end-of-support target for Java 11 and 17 builds, subject to extension by Microsoft. (Plan Your Java Upgrade Ahead of Microsoft’s 2027 Support Milestones)
- 12 November 2027: Microsoft Purview Data Governance: Support for Purview Data Governance ends. (Purview Data Governance is retiring: Your real asset isn’t the catalog, it’s the context)
- September 2028: Microsoft Build of OpenJDK: Earliest end-of-support target for Java 21 builds, subject to extension by Microsoft. (Plan Your Java Upgrade Ahead of Microsoft’s 2027 Support Milestones)
- September 2030: Microsoft Build of OpenJDK: Earliest end-of-support target for Java 25 builds, subject to extension by Microsoft. (Plan Your Java Upgrade Ahead of Microsoft’s 2027 Support Milestones)
Coming up
- 20 November 2026: Microsoft Windows: Aion Instruct is planned to begin rolling out to Windows Insider Preview devices. (Microsoft Aion Instruct: Advancing local AI on Windows and expanding to GPU)
- January 2027: Microsoft Windows: Aion Instruct is planned to roll out to retail Windows devices, completing the transition from Phi Silica. (Microsoft Aion Instruct: Advancing local AI on Windows and expanding to GPU)