E L S · T O O L S · G U A R D R A I L S Microsoft Foundry DECODED Enterprise AI Architecture for .NET Developers .NET C# Azure AI Agents MCP Foundry KÇ Kardel Rüveyda Çetin Digital Cloud Solution Architect · Microsoft D O T N E T C O N F 2 6 Sheraton Grand İstanbul 7 May 2026
Çetin Digital Cloud Solution Architect Microsoft · Türkiye Former Microsoft MVP Community Speaker X @kardelanite in kardelruveydacetin E D U C A T I O N 2013–2018 Y i l d i z T e c h n i c a l U n i v e r s i t y Mathematical Engineering Bahcelor 2019–2020 Y i l d i z T e c h n i c a l U n i v e r s i t y Computer Engineering / IT Master Degree (Non-Thesis) 2023–2025 Y i l d i z T e c h n i c a l U n i v e r s i t y Mathematical Engineering Master Degree (Thesis) E X P E R I E N C E 2016–17 Mavi · Overtech Internships 2018 Iconec Wave Jr. Frontend Dev 2018–21 Kariyer.net Software Specialist 2021–25 Dogus Tecnology SW Spec → Dev Lead 2026– Microsoft Digital CSA
E F O R E F R O N T AI adoption is accelerating 1.3B AI agents projected by 2028 S O U R C E · I D C 82% of organizations intend to integrate within 1–3 years S O U R C E · C A P G E M I N I 90% of enterprise software engineers will use AI code assistants by 2028 S O U R C E · G A R T N E R Technology · Budget · Talent — none of these are the biggest blocker. The blocker is TRUST. Dotnet Conf 26 K A R D E L R Ü V E Y D A Ç E T İ N
R E A L I T Y From one agent to many T O D A Y A T D 1 agent · 1 use case · controlled environment Manageable. Findable. Fixable. → B Y 2 0 2 8 A T D A T A A T D A T D A D A T Dozens of agents · Parallel tools · Cross-system reach Agent — Blue Tool — Purple Data — Cyan Dotnet Conf 26 K A R D E L R Ü V E Y D A Ç E T İ N
growing AI trust gap 41 31 52 44 49 62 0 10 20 30 40 50 60 70 80 2022 2024 Willing to rely Perceive as trustworthy Worried Why is trust shrinking? Power grows → risk grows. A wrong chatbot answer is annoying. An agent leaking customer data, misconfiguring a critical system, or breaching a legal boundary — is a different category entirely. Pilots stall. Investments don't scale. Dotnet Conf 26 K A R D E L R Ü V E Y D A Ç E T İ N
- 2 0 2 4 Azure AI Studio Try models. Write prompts. I g n i t e 2 0 2 4 Azure AI Foundry Build applications. I g n i t e 2 0 2 5 Microsoft Foundry Build, scale & integrate agents into the entire ecosystem. ai.azure.com — the URL never changed Each rename reflected expanding scope. The 'Azure' drop is intentional — Foundry now lives in Teams, M365, Copilot. Everywhere agents live. Dotnet Conf 26 K A R D E L R Ü V E Y D A Ç E T İ N
I N H E R I T E D Microsoft Foundry as an Azure Resource A Z U R E S U B S C R I P T I O N R E S O U R C E G R O U P Microsoft Foundry Azure resource · same RBAC · same policies · same cost view Two paths in ① ai.azure.com — direct, fastest entry ② Azure Portal → search 'Microsoft Foundry' Click 'Use with Foundry' — redirects to the Foundry portal. Azure infra power + Foundry's AI-focused UX.
E L R Ü V E Y D A Ç E T İ N S E T T I N G U P Y O U R L O C A L E N V I R O N M E N T Before you code VS Code / Visual Studio Editor + AI Toolkit extension Python 3.10+ Conda or venv environment .NET 9+ C# SDK for agents Azure CLI az login · DefaultAzureCredential
O N E Connecting to Foundry: the Project Client DefaultAzureCredential — Azure already knows who you are. F o u n d r y C l i e n t . c s — w r i t t e n o n c e using Azure.Identity; using Azure.AI.Projects; public static class FoundryClient { private const string Endpoint = "https://your-foundry.services.ai.azure.com/api/projects/proj"; public static AIProjectClient Create() => new AIProjectClient( new Uri(Endpoint), new DefaultAzureCredential()); } P r o g r a m . c s — r e - u s e d e v e r y w h e r e using Azure.AI.Projects; var project = FoundryClient.Create(); // Access agents, models, datasets, deployments… var agents = project.GetAgentsClient(); var agent = await agents.CreateAgentAsync( model: "gpt-5", instructions: "You are a helpful HR agent."); // No keys. No connection strings. Identity-driven.
F O U N D R Y One toggle, a new experience C L A S S I C • Azure OpenAI resources • Existing hubs & projects • Old Prompt Flow • AutoML & ML Pipelines • HuggingFace managed compute N E W F O U N D R Y • Discover · Build · Operate menu • Agent Builder · Workflows • Knowledge & Foundry IQ • Guardrails · Evaluations · Red Teaming • Memory · MCP server · 1,400+ tools Recommendation: keep old projects where they are, but flip the New Foundry toggle and start there. Dotnet Conf 26 K A R D E L R Ü V E Y D A Ç E T İ N
E W — D E E P D I V E Resource architecture & project hierarchy A S P E C T C L A S S I C ( H U B - B A S E D ) N E W ( F O U N D R Y P R O J E C T ) Resource Provider Microsoft.MachineLearningServices Microsoft.CognitiveServices Parent Resource AI Hub Foundry Resource (AIServices) Required Sibling Resources Storage Account, Key Vault None required by default Project Isolation Via Hub RBAC Native project-level RBAC Agent Service GA Preview only General Availability Foundry SDK & API Limited Full support ML Training AutoML, Pipelines Use hub-based project Prompt Flow Deprecated · sunset Jan 2027 Replaced by Framework Workflows Managed Compute (HF) Yes No Greenfield → New Foundry. Prompt Flow → migrate to Framework Workflows by Jan 2027. AutoML/HF → Classic. Dotnet Conf 26 K A R D E L R Ü V E Y D A Ç E T İ N
T W O O W N E R S Foundry — clear control vs data plane separation C O N T R O L P L A N E · I T A D M I N S Foundry Resource — top-level governance Create deployments Configure networking Manage projects Set encryption · Bicep · Azure Policy D A T A P L A N E · D E V E L O P E R S Foundry Project — child resource workspace Build agents Run evaluations Upload files Test in playground Inspect traces Optimize prompts Build once, govern centrally · scale freely.
T , W H E R E Layer responsibilities C O N T R O L P L A N E Foundry Resource (top-level) Owner: IT Admins ARM Control Plane API (consolidated) End-to-end Private Networking GA Managed VNET firewall & flow logs Azure Policy · RBAC · CMK encryption Bicep templates · Subscription mgmt D A T A P L A N E Foundry Project (child) Owner: Developers Build agents · Workflows · Memory Run evaluations & continuous monitoring Upload files · datasets · embeddings Test in playground · prompt optimize Inspect traces · debug · iterate Dotnet Conf 26 K A R D E L R Ü V E Y D A Ç E T İ N What is Azure Resource Manager?
C O N S I S T E N T · U N S C A L A B L E Classic — Agent Service in Preview D E L A Y E D Preview only · no SLA · compliance teams blocked production rollouts for weeks/months I N C O N S I S T E N T Single-agent only · limited tools · basic observability · debugging meant log-diving U N S C A L A B L E Connection strings everywhere · manual rotation · Key Vault sprawl · weak managed identity These three words are why portal UX was rewritten. Next slide → how each one was solved. Dotnet Conf 26 K A R D E L R Ü V E Y D A Ç E T İ N
M A N T I C K E R N E L → A G E N T F R A M E W O R K Multi-Agent Orchestration & Workflows Sequential A → B → C Deterministic. Predictable. Easy to debug. Intake → Analyze → Approve Group Chat Dynamic speaker selection. Planner · Researcher · Critic. AutoGen's original idea. Planner ⇄ Researcher ⇄ Critic Human-in-the- loop Pause for human approval. Financial · compliance · legal. Critical-decision gating. Agent → Human → Resume Visual builder YAML C# / Python / JS / Java SDK Versioned saves All three views stay in sync. Power Fx for conditional logic. Dotnet Conf 26 K A R D E L R Ü V E Y D A Ç E T İ N AutoGen Microsoft Research, 2023
R A B L E F U N C T I O N S + A G E N T F R A M E W O R K + S I G N A L R Durable Agent Orchestration 0 1 WORK Agent does the heavy lifting → 0 2 PAUSE wait_for_external_event → 0 3 PERSIST State on Azure · survives restarts · waits days → 0 4 STREAM SignalR → real-time UX → 0 5 RESUME Human approves · agent continues U S E C A S E S Incident Response Agent detects → root cause → SRE approves → remediation runs Provisioning Terraform diff prepared → platform team reviews → applied Document Review Draft generated → legal/compliance signs → published Continuous Evaluation runs in the background. Tracing GA shows every step. Nothing is lost. Dotnet Conf 26 K A R D E L R Ü V E Y D A Ç E T İ N Durable Functions Microsoft Documentation
N T T A X O N O M Y Three agent types — pick the right one Prompt Agent kind: prompt Declarative: model + instructions + tools + knowledge. No code. U s e w h e n : Customer support · doc summarization · RAG Q&A · the majority of production agents. Hosted Agent kind: hosted Custom code in container. Foundry hosts and manages. U s e w h e n : LangGraph · custom frameworks · multi-stage state machines · domain-specific coordination. Workflow YAML-based Orchestrates multiple agents using Sequential / Group Chat / HITL. U s e w h e n : Marketing Analyst → Copywriter → Editor. Preview — opt in via allow_preview=True. Workflow runs the process. Prompt + Hosted Agents do the work inside it. Dotnet Conf 26 K A R D E L R Ü V E Y D A Ç E T İ N AI Agent Orchestration Patterns
G R A T I O N · B U T A L S O A N O P P O R T U N I T Y Classic agents don't auto-show in New Foundry Heads up Agents created in Classic are not automatically visible in the New Foundry experience. To work in the new environment, re-register them as Prompt Agent or Workflow. But also — an opportunity The new runtime is different. Re-registering forces a review. Most teams discover that half of their Classic agents can be simplified during this migration. Cleaner instructions · fewer tools · better governance. Dotnet Conf 26 K A R D E L R Ü V E Y D A Ç E T İ N
A U G M E N T E D G E N E R A T I O N · L E W I S E T A L . 2 0 2 0 Basic RAG Q U E R Y → E M B E D → V E C T O R D B → T O P - K → L L M → A N S W E R How it works Embed the question. Search a vector store for similar passages. Inject top-k results into the prompt. Model answers with citations. Strengths · Limits ✓ Cheap · fast · easy to start. ✓ Works for one document, single source. ✗ Struggles with multi-hop queries. ✗ No reranking · no agentic decomposition. Foundation pattern. Everything that follows is layered on this. Dotnet Conf 26 K A R D E L R Ü V E Y D A Ç E T İ N RAG Microsoft Learn
E S E A R C H · C H A I N E D & A D A P T I V E R E T R I E V A L Advanced RAG · CoRAG Q U E S T I O N Q U E R Y A N A L Y S I S W E B S E A R C H V E C T O R S T O R E F O U N D R Y I Q R E R A N K S Y N T H E S I Z E S E L F - C H E C K R E - P L A N ↻ I T E R A T E Multi-hop reasoning Query decomposition Cross-source retrieval Self-correction loop Microsoft Research's CoRAG paper: agents as first-class coordinators in retrieval pipelines. Dotnet Conf 26 K A R D E L R Ü V E Y D A Ç E T İ N
I O R , N O T J U S T K N O W L E D G E Fine-tuning — three techniques S T A R T H E R E Supervised FT Show example conversations. Model learns style, tone, format. E x a m p l e Always reply in bullet points, formal tone. A D V A N C E D Reinforcement FT No labels needed. Define a grader. Model learns to maximize score. E x a m p l e Complex math · multi-step reasoning tasks. P R E F E R E N C E T U N I N G DPO A/B preference data. Model learns the preferred style. E x a m p l e Customer feedback shows users like A over B. Fine-tuning + RAG Behavior + Knowledge — complementary, not competing. March 2026 additions Dedicated fine-tuning CLI · cost projection · Fireworks AI BYO Weights (preview). Dotnet Conf 26 K A R D E L R Ü V E Y D A Ç E T İ N Fine-Tuning Microsoft Learn
' T O ' Y O U R E N T I R E B U I L D I N G ' Foundry IQ — managed knowledge for advanced RAG F O U N D R Y I Q Permission-aware Azure Blob auto-chunk · embed · index SharePoint auto-chunk · embed · index OneLake auto-chunk · embed · index Public Web auto-chunk · embed · index Azure AI Search auto-chunk · embed · index Fabric Data auto-chunk · embed · index Agentic retrieval Decomposition Purview labels Private networking Service-level CMK Dotnet Conf 26 K A R D E L R Ü V E Y D A Ç E T İ N Foundry IQ Learning Path
E + C H A T S U M M A R Y Memory — agents that remember you User Profile Memory Preferences · language · past decisions. Persists across sessions. User prefers concise replies. Speaks Turkish. Works in HR. Chat Summary Memory Distilled summary of every conversation. Carried into the next session. Last week: discussed parental leave policy with Ahmet. P I P E L I N E E X T R A C T → C O N S O L I D A T E → R E T R I E V E Limits: 10,000 memories per scope · 100 scopes per store. Native API in SDK 2.0.0 stable. Dotnet Conf 26 K A R D E L R Ü V E Y D A Ç E T İ N Memory in Microsoft Foundry Agent Service
C P · C U S T O M · P R I V A T E Expanded Tool Catalog · 1,400+ tools 1,400 + tools available vs. 4 in Classic C L A S S I C H A D Azure AI Search · File Search · Code Interpreter · Custom funcs Built-in Bing · Code Interpreter · Image Gen · Computer Use · SharePoint · Fabric · Browser MCP Servers Remote · Local · 4 auth methods (Key, Entra Agent, Managed Identity, OAuth Passthrough) Custom OpenAPI 3.0 spec · A2A endpoints · any REST API as a tool Private Catalog Org-only tools · proprietary IP · team-scoped visibility OAuth Identity Passthrough · agent acts on behalf of the user, not as itself. Dotnet Conf 26 K A R D E L R Ü V E Y D A Ç E T İ N Ushering in the next era of agentic AI with tools in Microsoft Foundry
G · A G E N T I N S P E C T O R · I N S P I R E M E AI Toolkit for VS Code v0.30.0 agent.py — VS Code A I T O O L K I T Models Tools Agent Builder Inspector GitHub Copilot Tool Catalog 1,400+ tools · MCP · Built-in · Custom · Private — without leaving the editor. Agent Inspector F5 debug · breakpoints · variable inspection · streaming visualization. Inspire Me Natural language → agent instructions. Great starter for prompt newcomers. Local development, production parity. "Worked locally, broke in prod" — finally history. Dotnet Conf 26 K A R D E L R Ü V E Y D A Ç E T İ N
T E A M I N G · C O N T I N U O U S E V A L U A T I O N G A Integration · Test · Evaluation Out-of-the-box Coherence · Relevance · Groundedness · Retrieval · Safety. Zero setup. Custom Your business rules · tone · domain compliance — codified. Continuous Sample production traffic · evaluate · alert on drift via Azure Monitor. Evaluation · 18 metrics Synthetic dataset → score → diagnose → fix → re-evaluate. ✓ Eval results link directly to agent traces ✓ Prompt Optimizer integrated into Playground ✓ Same evaluators in CI/CD, local, hosted runs Red Teaming Adversarial attack simulation. Find weaknesses before customers do. Jailbreak attempts PII exfiltration probes Out-of-scope behavior Tool-call · response guardrails Quality is no longer a release-day checkbox — it's a live signal. Dotnet Conf 26 K A R D E L R Ü V E Y D A Ç E T İ N
· B I Z C H A T · C O N T A I N E R · M C P / A 2 A Publishing — meet users where they live Microsoft 365 Teams · M365 · BizChat — one click Container Anywhere: on-prem · cloud · hybrid · 6 new Azure regions Open Protocols MCP + A2A · 4 auth methods · OAuth Passthrough AI Gateway Azure API Management — rate limits · quotas · monitoring Azure Policy Built-in governance · ARM control plane · private networking GA Delayed → GA · Inconsistent → multi-agent orchestration · Unscalable → enterprise-ready Dotnet Conf 26 K A R D E L R Ü V E Y D A Ç E T İ N Publish agents to Microsoft 365 Copilot and Microsoft Teams
I · A N T H R O P I C · M E T A · M I S T R A L · x A I · D E E P S E E K · C O H E R E · M I C R O S O F T · N V I D I A · O P E N S O U R C E 11,000+ frontier & open models — one platform 11,000 + models in catalog AWS Bedrock & Google Vertex are siloed. Foundry isn't. Azure OpenAI GPT-5.4 · GPT-5.4 Pro · Sora · o4-mini Anthropic Claude Opus 4.6 · Sonnet 4.6 · Haiku 4.5 Meta Llama 4 Scout · Maverick · Llama-Guard xAI Grok 4 · 4.1 Fast · 4.2 · Code Fast 1 DeepSeek DeepSeek-R1 · V3.1 · MAI-DS-R1 Mistral Mistral Large 3 · Mixtral · OCR · Document AI Microsoft Phi-4 · Phi-4 Reasoning Vision · Aurora Open + NVIDIA Llama · Qwen · Nemotron · DBRX · Snowflake Switch models without rewriting. Same endpoint, same governance — only the model changes. Dotnet Conf 26 K A R D E L R Ü V E Y D A Ç E T İ N
Ü V E Y D A Ç E T İ N L L M · R E A S O N I N G · M U L T I M O D A L · C O D I N G Model types — pick by use case LLMs / SLMs General-purpose: conversation, summarization, Q&A. W H E N T O U S E Enterprise copilots · customer support · knowledge search. Reasoning Logic · multi-step reasoning · complex problem solving. W H E N T O U S E Financial analysis · scientific workflows · decision support. Multimodal Process images, video, documents alongside text. W H E N T O U S E Marketing content · computer vision · doc automation. Coding Tuned for software development. W H E N T O U S E Code gen · bug fixing · refactoring · developer copilots. o3 · 20% fewer critical errors. o4-mini · cost & speed optimized for high volume.
Ü V E Y D A Ç E T İ N I M A G E · D O C U M E N T · C O D E · F R O N T I E R Specialized model categories Image / Video Sora GPT-Image-1 FLUX.1 Kontext Pro FLUX 1.1 Pro OCR / Document / Embed Mistral Document AI Embed v4 / Rerank 4 Coding Grok Code Fast 1 GPT-5 Codex GPT-5.1 Codex GPT-5.1 Codex Mini GPT -5.1 / 4 Family GPT-5.1* GPT-4* Llama 3.3 Llama-4-Maverick Grok 3* Grok 4 / 4.1 DeepSeek-V3* Mistral Large 3 Command A Reasoning Frontier GPT-5-Pro DeepSeek-R1* MAI-DS-R1 Grok 3 / Grok 3 mini Grok 4 / 4.1 Fast GPT-4.1 family — fastest growing usage. Sora delivered enterprise-first.
Ü V E Y D A Ç E T İ N W H E R E P R O C E S S E D · H O W B I L L E D · W H E R E S T O R E D How will you deploy? Global Standard Any region · pay-per-token · most flexible Global Provisioned Reserved PTU · guaranteed performance Global Batch 50% off · 24-hr turnaround · bulk async Data Zone EU / US data residency · GDPR friendly Developer Eval only · no SLA · for fine-tune testing Priority Processing PREVIEW · dedicated lane · between Standard & Provisioned · for latency-critical P O R T A L C A P A B I L I T Y Azure OpenAI · DeepSeek · xAI Native in New Marketplace Native in New HuggingFace managed compute Classic only Model Catalog Requires project context in New
Ü V E Y D A Ç E T İ N G A · A Z U R E O P E N A I GPT-5.4 Mini Lightweight High-volume · low-complexity workloads Routing-friendly Pairs with GPT-5.4 in routed deployments Reliable output Better instruction following · solid JSON $0.75 / $4.50 input · output · per million tokens Designed to handle the 80% — classification, extraction, simple tool calls — and route only the 20% of reasoning-heavy work to standard 5.4. GPT-5.4 thinks. GPT-5.4 Mini runs.
Ü V E Y D A Ç E T İ N G A · A Z U R E O P E N A I · P R O D U C T I O N - R E A D Y GPT-5.4 Production-ready Long-horizon agentic workflows Agentic execution Tools · files · agent handoffs Computer Use Built-in screen interaction $2.50 / $15.00 input · output · per million tokens A different category for production. Holds intent across long, multi-step workflows. Lower latency for real-time. Available in Global Standard and Data Zone US.
Ü V E Y D A Ç E T İ N G A · D E E P A N A L Y T I C S GPT-5.4 Pro Deep analysis Scientific & analytical workloads Multi-path reasoning Explores multiple solution paths Research-grade Accuracy over speed $30.00 / $180.00 input · output · per million tokens When precision matters more than latency. Scientific research, complex decision support, multi-hypothesis analysis. The premium reasoning tier.
Ü V E Y D A Ç E T İ N G A · A N T H R O P I C · F I R S T - P A R T Y Claude Opus 4.6 1M token context ≈ 2,000 pages or a mid-size repo Adaptive thinking low · medium · high · max effort Context compaction Auto-summarizes long agent sessions $5 / $25 input · output · per million tokens Anthropic's most capable reasoning model — first-party in Foundry. 128K output token capacity for end-to-end refactors, comprehensive analysis, complete documentation in a single response.
Ü V E Y D A Ç E T İ N G A · A N T H R O P I C · C O S T - E F F I C I E N T S C A L E Claude Sonnet 4.6 1M token context Same capacity as Opus Cost-efficient Opus quality at scale-friendly price Agentic workflows Coding · ops · content generation Lower cost same 1M context · same 128K output Opus for accuracy. Sonnet to deliver that quality at scale. Optimized for coding, agentic workflows, and professional content generation.
Ü V E Y D A Ç E T İ N G A · x A I · F I R S T G A x A I M O D E L I N F O U N D R Y Grok 4.0 Complex reasoning Multi-step analysis · planning Frontier-tier Reached GA in Feb 2026 Foundry-deployable Serverless or provisioned $5.50 / $27.50 input · output · per million tokens xAI's frontier reasoning model — the first xAI model to reach GA on Foundry.
Ü V E Y D A Ç E T İ N P R E V I E W · x A I · H I G H - V O L U M E Grok 4.1 Fast Non-reasoning variant Speed and throughput first High-volume Classification · extraction · routing Cost-leader Among the cheapest on Foundry $0.20 / $0.50 input · output · per million tokens Pair with Grok 4.0: complex reasoning on 4.0, fast and cheap operations on 4.1 Fast. Same family, different tiers.
Ü V E Y D A Ç E T İ N G A · x A I · C H A T M O D E L Grok 4.2 Beta → GA Production-ready since March 30 Conversational Less corporate · more flexible Brand voice Distinctive tone for unique products Chat-tuned Serverless or provisioned throughput Production-grade conversational model with personality. For customer- facing chat where a distinct brand voice matters.
Ü V E Y D A Ç E T İ N P R E V I E W · B L A C K F O R E S T L A B S · T E X T - H E A V Y I M A G E R Y FLUX.2 Flex Text-heavy images Typography renders correctly UI prototyping Mockups · infographics · marketing Multi-prompt fidelity High instruction adherence $0.05 per megapixel · budget-friendly When the text inside the image must be correct. Reports → infographics, marketing collateral, UI mocks. The pipeline picks FLUX.2 Flex when text matters.
Ü V E Y D A Ç E T İ N G A · M I C R O S O F T · S M A L L · M U L T I M O D A L · R E A S O N I N G Phi-4 Reasoning Vision 15B Small but mighty 15B parameters · big-model behavior Multimodal Charts · diagrams · documents · forms Chain-of-thought Step-by-step visual reasoning Edge-friendly On-prem · containerized · constrained envs GPT-4 Vision-class capability in a much smaller, cheaper model. Ideal where you can't run frontier models — on-prem, edge, regulated environments.
R I E S S U P P O R T E D Model Router · one endpoint, the right model O N E E N D P O I N T /router S I M P L E Cheap model GPT-5.4 Mini · Grok 4.1 Fast S T A N D A R D Balanced model GPT-5.4 · Sonnet 4.6 C O M P L E X Quality model Opus 4.6 · GPT-5.4 Pro BALANCED COST QUALITY Dotnet Conf 26 K A R D E L R Ü V E Y D A Ç E T İ N
T A B · N O T F I V E D A S H B O A R D S Observability & Monitoring C L A S S I C ✗ Manual Application Insights setup ✗ SDK-based manual tracing ✗ No agent dashboard ! Resource-level metrics only ✗ Five tools to debug one agent N E W F O U N D R Y ✓ Built-in Application Insights ✓ Real-time conversation tracing — GA ✓ Agent monitoring dashboard ✓ Resource + Project level metrics ✓ Multi-cloud agent registration ✓ Continuous Evaluation → Azure Monitor ✓ Notification Center · email alerts ✓ Managed VNET firewall & flow logs Quality is now a metric — alert on drift, trend it, monitor it like CPU. Dotnet Conf 26 K A R D E L R Ü V E Y D A Ç E T İ N
L A N G U A G E S S T A B L E 2 . 0 . 0 Preview → RC → GA — platform is mature F e b 2 0 2 6 REST API v1 GA Chat · Responses · Embeddings · Fine-tune · Vector Stores M a r 6 Python 2.0.0 stable azure-ai-projects — first stable, openai bundled M a r 6 JS/TS 2.0.0 stable @azure/ai-projects — first stable M a r 2 7 Java 2.0.0 azure-ai-projects — first Java GA A p r 1 .NET 2.0.0 Final piece — all four languages on the same REST surface $ pip install azure-ai-projects ← one package: agents · evals · memory · inference · openai · azure-identity Dotnet Conf 26 K A R D E L R Ü V E Y D A Ç E T İ N
G O · B R O W N F I E L D C H E C K Move to New Foundry when… G O T O N E W F O U N D R Y Building agentic applications Multi-agent workflows needed Tool Catalog · Memory · Foundry IQ Publish to M365 · Teams · BizChat Starting greenfield Fine-tuning · evaluation · red teaming S T A Y O N C L A S S I C I F … Prompt Flow dependency — sunset Jan 2027 → migrate to Framework Workflows HuggingFace managed compute models AutoML / ML Pipelines workloads Full end-to-end network isolation (or use SDK / CLI in New) Direction is clear: New Foundry is the future. Start there. Dotnet Conf 26 K A R D E L R Ü V E Y D A Ç E T İ N
E I 2 0 2 6 The business case · 327% ROI in 3 years 327% Return on Investment over 3 years <6 mo Payback period fastest in category $37.9M Net Present Value for the composite org Source: Forrester Total Economic Impact study — Microsoft Foundry, February 2026. Dotnet Conf 26 K A R D E L R Ü V E Y D A Ç E T İ N
Ü V E Y D A Ç E T İ N B I G G E S T V A L U E D R I V E R · N O T W H A T F O R R E S T E R E X P E C T E D $15.7M · developer productivity gain $15.7M over 3 years Senior engineers spent ⅓ of their time gluing tools, rebuilding pipelines, and managing bespoke governance. 30-40% less development time +35% technical team productivity 32% of orgs retired legacy AI tools 67% moved for security & governance Trust isn't just a feature. It's the precondition for ROI at scale.
kardelruveyda
maroon1st
simas
terisuke
dip_tech
kishikawakatsumi
o0h
makenowjust
yoshikikasama
soutaro
koher
ryotanakaya
m_seki
leimatthew05
honnibal
soudai
lauravandoore
samtorres
swwweet
rakko
orderedlist
aleyda
.jpg){kind=avatar}
cargoodrich
iamctodd
asonas
