Browse Artificial Intelligence Community (139)

kinfey breaks down what changes when an AI agent can run tools and create side effects, then shows how to put real execution boundaries around those tools using OpenSandbox on AKS with Kata pod VMs, plus a comparison to Azure Container Apps Sandboxes and Dynamic Sessions.

Sonia Cuff’s September 2026 FinOps roundup pulls together cost and pricing updates across Microsoft’s AI and cloud stack, including Azure AI Foundry and Azure OpenAI pricing changes, new model options, and GitHub Copilot billing/usage metrics updates, plus a few Azure cost-optimization and purchasing notes.

simonjj explains when “decision models” are a better fit than general-purpose LLM calls, then benchmarks a self-hosted Winnow model served by Ollaya on Azure Container Apps serverless GPU against GPT-5.4 Nano via Azure OpenAI, focusing on latency, throughput, accuracy, and token usage trade-offs.

frrey explains how Azure ExpressRoute and Microsoft’s AI Points of Presence (AI PoPs) are used to connect distributed, third-party GPU capacity to Azure with private networking, aiming to keep a consistent Azure operating model even when compute runs outside hyperscale regions.

Rebekah Midkiff lays out a reference architecture for governing AI agents across multiple Microsoft Entra tenants, combining Agent 365, Entra Agent ID, Microsoft Foundry, Copilot Studio, Purview, and an Azure API Management AI gateway to enforce identity, policy, tool access, and observability at runtime without collapsing tenant boundaries.

rickliev outlines Microsoft’s message for TM Forum Innovate Americas 2026: scaling agentic AI in telecom depends on trusted, governed data plus shared operational context. The post connects “agent-ready” data concepts (meaning, provenance, permissions) with Microsoft Fabric capabilities like OneLake, Fabric IQ, and Real-Time Intelligence.

kinfey describes a hands-on experiment comparing a traditional GitHub Copilot tool-calling agent loop with a Jev-based “decision harness” that produces typed choices and confidence signals before execution. The post focuses on splitting responsibilities between models and ordinary code to reduce unnecessary model work and improve observability.

Christine Caggiano explains how a multidisciplinary team used the Microsoft Discovery app to design an inspectable, reproducible bioinformatics workflow for phage therapy—connecting curated literature, genomic analysis, and external tools via MCP-based adapters, with validation steps and safety checks built into each stage.

paulyuk introduces the Azure Functions Hosted Skills canvas inside the GitHub Copilot app, showing how to build and test an event-driven Function App locally, invoke triggers on demand, and debug runs with logs and evidence links. The post also covers deployment to Azure with azd and exposing skills as MCP tools.

Arun Kumar Thiagarajan explains how to connect Microsoft Playwright Workspaces (Playwright Cloud Browsers) to agent applications via Model Context Protocol (MCP), using the GitHub Copilot app as the example client. The post focuses on provisioning the Azure workspace, wiring up the remote MCP endpoint, and running safer human-approved browser workflows with evidence and diagnostics.

Robin_Lester introduces TypeSafe AI’s Jev “System One” model and shows how to use it as a structured judgement layer inside Microsoft Foundry agent workflows, including an OpenAPI tool setup and an example triage flow that returns typed decisions with probabilities and confidence.

Wiliam Rosa explains how to use Azure Databricks ai_query to run a “decision model” (closed-set classification with probabilities) directly from SQL against Unity Catalog-governed tables, avoiding chat-style outputs and extra orchestration. The post covers required warehouse/runtime constraints, structured return types, hosting options, and governance caveats in Unity Gateway.

emiller635 compares three practical ways to evaluate and observe AI agents: self-hosted Langfuse on an Azure VM, a Microsoft Foundry hosted agent, and a standalone agent on Azure Container Apps. The post breaks down what each option gives you for tracing, scoring, and dashboards, and what you still have to build and operate yourself.

Wiliam_Rosa breaks down a practical rollout pattern for testing newly released “frontier” AI models without letting costs spike: give immediate experimental access, but isolate spend risk using layered budgets in Azure Databricks Unity AI Gateway, then decide promotion or rollback using benchmarks, user feedback, and OpenTelemetry cost traces.

Wiliam_Rosa explains how Azure Databricks’ Unity Gateway CLI (ug) centralizes authentication, routing, governance, and usage tracking for multiple coding agents (Claude Code, Codex, Gemini, and GitHub Copilot CLI) by reusing Unity Catalog permissions, rate limits, and budgets instead of scattering provider keys across developer machines.

stevenbucher announces the Azure Policy Linter as an open-source CLI and .NET library, aimed at catching common Azure Policy definition mistakes before policies are assigned. The post explains how to install and run it, how rule sets and JSON output work, and highlights several high-impact rules from the catalog.

kinfey explains how ONNX Runtime Mobius builds generative models directly as ONNX graphs, why that matters for tricky architectures, and how the JevONNX sample converts a Jev-style “typed decision” workflow from a GGUF model into a local CPU ONNX package with reproducible pins and quantization reports.

Matt Kazanowsky describes an Azure reference architecture that turns assembly videos into illustrated manufacturing work instructions using multimodal GenAI. It breaks down the end-to-end pipeline (ingestion, transcription, human review gates, step extraction, frame selection, and document generation) and calls out practical security controls for storage, identity, networking, and AI safety.

puneetguptams shares practical configuration guidance for Azure SRE Agent, based on patterns observed in production telemetry. The post focuses on connecting the right data sources, routing incidents to the right specialist via response plans, writing useful skills, deciding when to require approvals, and ensuring the agent has the permissions needed to complete the last step.

Lee Stott describes a developer demo that shows how to let an AI agent produce a decision brief while keeping real-world actions (like quarantining inventory) under deterministic, authenticated, auditable application control using MCP, Microsoft Foundry Hosted Agents, Entra auth, and Azure Blob Storage concurrency safeguards.

nitya invites developers to a Foundry Friday AMA focused on improving production AI agents using Azure AI Foundry’s Insights and Agent Optimizer, with a live demo and Q&A on turning trace patterns into better instructions, tools, skills, and model choices.

abbyhamilton highlights several Catalyst series customer stories showing how organizations use Microsoft Azure AI infrastructure alongside NVIDIA to run high-performance AI workloads, including real-time weather forecasting, multimodal health screening, large-scale biological databases, photorealistic digital twins, and rare-disease patient identification.

Simon Frost explains how he used the Microsoft Discovery app to speed up exploratory analysis of a large wastewater metagenomics dataset, while keeping human scientific judgment in the loop. The post walks through agent-assisted data wrangling, API enrichment, phylogenetics, and visualization to produce auditable, reproducible findings.

Lee Stott walks through Fibey Field Ops, a synthetic field-service assistant built with Microsoft Foundry Hosted Agents, MCP, and Azure Container Apps, focusing on the practical engineering work around tool contracts, identity boundaries, deployment sequencing, and debugging/observability rather than just model choice.

Nate_Waters summarizes a wave of Azure Kubernetes Service (AKS) releases aimed at running AI workloads more safely and efficiently, covering isolation features (pod sandboxing, confidential GPU), agent-focused platform components (GitOps, burst scaling, GPU profiling), and capacity improvements for large-scale clusters and hybrid/edge deployments.

lily-ma introduces Azure Functions managed connectors (public preview) and shows how connector triggers and typed connector clients let functions react to events and call actions across services like SharePoint and Teams without hand-rolling webhook registration or OAuth token handling.

Trinidad Salazar explains how Microsoft and Dapple are integrating topology-aware, purpose-built GPU clusters into an Azure-native operating model, using AKS as the control plane so teams can run large-scale training, fine-tuning, and inference while keeping familiar governance and operational workflows.

kinfey explains how Azure KARS (a Kubernetes reference stack) can run “coding-agent” CLIs as governed workloads, focusing on blast-radius control, credential isolation, and consistent policy enforcement across multiple agent runtimes like Copilot CLI, Claude Code, and Codex.

piyush05 explains how to deploy Elastic’s serverless Elasticsearch Vector Database through Azure Native Integrations, then use it for hybrid (keyword + vector) semantic search and RAG. The post also covers Azure portal onboarding, permissions, Entra ID SSO, and Azure-connected billing for Elastic resources.

kinfey lays out a practical way to evaluate text-to-image models using the HEIM benchmark, focusing on repeatable testing across quality, instruction following, safety, bias, robustness, multilingual prompts, and production constraints like latency and cost.

Wiliam_Rosa explains why Azure Databricks Genie Agents need the same ongoing evaluation discipline as production ML models, and walks through Genie Agent Benchmarks, including SQL-based “Chat mode” scoring, LLM-judged “Agent mode”, and a practical workflow for monitoring and improving agent accuracy over time.

Sayanika walks through building a conversational analytics agent that queries data in OneLake using Microsoft Fabric semantic models and Fabric IQ Ontology, with a Copilot Studio setup that normalizes messy location inputs (e.g., abbreviations and misspellings) before running ontology-backed queries.

Sayanika outlines an architecture pattern for a retail “growth intelligence” agent that uses Azure AI Foundry (Foundry IQ) to orchestrate reasoning across Microsoft Fabric (Fabric IQ + ontology), Dynamics 365 Finance & Operations via an ERP MCP server, SharePoint policy documents, and web intelligence, backed by Azure AI Search.

greenie-msft introduces Dynamic Workflows for Azure Functions hosted skills, showing how to run multi-step, long-running AI-driven plans using Durable Functions so work can continue beyond the original trigger while keeping state, retries, and timers durable.

fiveisprime previews a guided GitHub Copilot experience in VS Code that aims to take developers from an app idea to a deployed Azure application using a structured workflow with explicit checkpoints for scaffolding, local setup, and deployment.

Wiliam_Rosa breaks down how Azure Databricks’ internal “AI SRE” agent speeds up incident investigation by gathering context automatically before engineers start debugging. The post focuses on an operational pattern: deterministic health checks and version-controlled runbooks first, then LLM-based synthesis—plus the guardrails needed to keep observability access safe.

Gaurav Bhardwaj proposes a Microsoft Foundry + Azure AI Search architecture that preserves meaning when turning job descriptions into recruiter-reviewable Boolean searches, with structured extraction, terminology provenance, an explicit approval step, and evaluation/observability to catch errors like “preferred” skills becoming mandatory.

Nivas Narayanasamy lays out a practical architecture for making private Azure VMware Solution workloads reachable by Azure AI Foundry agents without moving data. The post shows how to use MCP “bridge” servers on VNet-integrated Azure Container Apps, with ExpressRoute, private endpoints, Entra managed identity, and Key Vault-backed secrets to keep access private and read-only.

Pamela_Fox shares the full set of MCP Live session recordings and resources, covering the Model Context Protocol’s latest spec updates, MCP server/client development in VS Code and GitHub Copilot, enterprise governance with Microsoft Foundry Toolboxes, and evolving MCP authentication patterns.

Satoshi Kubota proposes using a Data Center WordCode Layer (DCWL)—a fixed-length 32-bit word encoding originally aimed at Azure text-processing workloads—to speed up Copilot for Windows and other Windows client text-heavy pipelines such as local RAG, semantic search, and log analysis.