{"id":2084,"date":"2026-04-09T16:33:53","date_gmt":"2026-04-09T23:33:53","guid":{"rendered":"https:\/\/devblogs.microsoft.com\/foundry\/?p=2084"},"modified":"2026-04-09T16:33:53","modified_gmt":"2026-04-09T23:33:53","slug":"whats-new-in-microsoft-foundry-mar-2026","status":"publish","type":"post","link":"https:\/\/devblogs.microsoft.com\/foundry\/whats-new-in-microsoft-foundry-mar-2026\/","title":{"rendered":"What&#8217;s new in Microsoft Foundry | March 2026"},"content":{"rendered":"<h2>TL;DR<\/h2>\n<ul>\n<li><strong>Foundry Agent Service (GA):<\/strong> The next-gen agent runtime is production-ready \u2014 Responses API-based, end-to-end private networking, MCP auth expansion (including OAuth passthrough), Voice Live preview, and hosted agents in 6 new regions.<\/li>\n<li><strong>GPT-5.4 + GPT-5.4 Pro (GA):<\/strong> Production-grade reasoning with integrated computer use, stronger instruction adherence, and dependable multi-step execution. Standard at $2.50\/$15 per million tokens; Pro at $30\/$180 for deep analytical workloads.<\/li>\n<li><strong>GPT-5.4 Mini (GA):<\/strong> Cost-efficient small model for classification, extraction, and lightweight tool calls \u2014 the high-volume tier in a GPT-5.4 routing strategy.<\/li>\n<li><strong>Phi-4 Reasoning Vision 15B:<\/strong> Multimodal reasoning meets the Phi family \u2014 visual understanding with chain-of-thought for charts, diagrams, and document layouts.<\/li>\n<li><strong>Evaluations (GA) + Continuous Monitoring:<\/strong> Out-of-the-box and custom evaluators with continuous production monitoring piped into Azure Monitor \u2014 quality isn&#8217;t a pre-ship checkbox anymore, it&#8217;s a live signal.<\/li>\n<li><strong>azure-ai-projects SDK (GA):<\/strong> Python 2.0.0, JS\/TS 2.0.0, and Java 2.0.0 all shipped stable releases targeting the GA REST v1 surface. .NET 2.0.0 followed on April 1. The <code>azure-ai-agents<\/code> dependency is gone \u2014 everything lives under <code>AIProjectClient<\/code>.<\/li>\n<li><strong>Fireworks AI on Foundry (Preview):<\/strong> High-performance open model inference \u2014 DeepSeek V3.2, gpt-oss-120b, Kimi K2.5, and MiniMax M2.5 with bring-your-own-weights support.<\/li>\n<li><strong>NVIDIA Nemotron Models:<\/strong> Open NVIDIA models now first-class in the Foundry catalog, announced at GTC alongside the Agent Service GA.<\/li>\n<li><strong>Grok 4.2 (GA):<\/strong> xAI&#8217;s refreshed chat model graduates from beta.<\/li>\n<li><strong>Priority Processing (Preview):<\/strong> Dedicated compute lane for latency-sensitive AI workloads \u2014 reserved capacity for real-time agents and customer-facing chat.<\/li>\n<li><strong>Palo Alto Prisma AIRS + Zenity (GA):<\/strong> Third-party runtime security integrations for prompt injection, data leakage, and tool misuse detection.<\/li>\n<li><strong>Tracing (GA):<\/strong> End-to-end agent trace inspection with sort, filter, and data model refinements.<\/li>\n<li><strong>PromptFlow Deprecation:<\/strong> Migration to Microsoft Framework Workflows required by January 2027.<\/li>\n<\/ul>\n<div class=\"d-flex\"><a class=\"cta_button_link\" href=\"https:\/\/learn.microsoft.com\/azure\/foundry\/quickstarts\/get-started-code\" target=\"_blank\" rel=\"noopener\">Build Your First Agent with Foundry Agent Service<\/a><\/div>\n<h2>Join the community<\/h2>\n<p>Connect with 25,000+ developers on <a href=\"https:\/\/aka.ms\/foundry\/discord\">Discord<\/a>, ask questions in <a href=\"https:\/\/aka.ms\/foundry\/forum\">GitHub Discussions<\/a>, or <a href=\"https:\/\/devblogs.microsoft.com\/foundry\/category\/whats-new\/feed\/\">subscribe via RSS<\/a> to get this digest monthly.<\/p>\n<hr \/>\n<h2>Models<\/h2>\n<h3>GPT-5.4 + GPT-5.4 Pro (GA)<\/h3>\n<p><strong>GPT-5.4<\/strong> went generally available on March 5 \u2014 and this one is about reliability, not raw intelligence. If you&#8217;ve been fighting task drift, mid-workflow failures, and inconsistent tool calling in production agents, GPT-5.4 is designed specifically for those problems. Stronger reasoning over long interactions, better instruction adherence, and integrated computer use capabilities for structured orchestration of tools, files, and data extraction.<\/p>\n<p><strong>GPT-5.4 Pro<\/strong> is the premium variant for when analytical depth matters more than latency \u2014 multi-path reasoning evaluation, improved stability across long reasoning chains, and enhanced decision support for scientific research and complex trade-off analysis.<\/p>\n<table>\n<thead>\n<tr>\n<th>Model<\/th>\n<th>Context<\/th>\n<th>Pricing (per M tokens)<\/th>\n<th>Best For<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><strong>GPT-5.4<\/strong> (\u2264272K input)<\/td>\n<td>272K<\/td>\n<td>$2.50 input \/ $0.25 cached \/ $15 output<\/td>\n<td>Production agents, coding, document workflows<\/td>\n<\/tr>\n<tr>\n<td><strong>GPT-5.4<\/strong> (&gt;272K input)<\/td>\n<td>Extended<\/td>\n<td>$5.00 input \/ $0.50 cached \/ $22.50 output<\/td>\n<td>Large-context reasoning<\/td>\n<\/tr>\n<tr>\n<td><strong>GPT-5.4 Pro<\/strong><\/td>\n<td>Full<\/td>\n<td>$30 input \/ $180 output<\/td>\n<td>Deep analysis, scientific reasoning<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>Deployment: Standard Global and Standard Data Zone (US) at launch, with additional options coming.<\/p>\n<blockquote><p><strong>Action:<\/strong> Deploy GPT-5.4 from the <a href=\"https:\/\/ai.azure.com\/catalog\">model catalog<\/a>. If you&#8217;re running GPT-5.2 in production, GPT-5.4 is a drop-in upgrade with better instruction following and fewer mid-workflow failures.<\/p><\/blockquote>\n<div class=\"d-flex\"><a class=\"cta_button_link btn-secondary\" href=\"https:\/\/techcommunity.microsoft.com\/blog\/azure-ai-foundry-blog\/introducing-gpt-5-4-in-microsoft-foundry\/4499785\" target=\"_blank\" rel=\"noopener\">Read the GPT-5.4 Announcement<\/a><\/div>\n<h3>GPT-5.4 Mini (GA)<\/h3>\n<p><strong>GPT-5.4 Mini<\/strong> shipped on March 17 \u2014 OpenAI&#8217;s latest small model optimized for fast, cost-efficient tasks like classification, extraction, and lightweight tool calls. A step up from GPT-5 mini in instruction following and structured output reliability, at a fraction of the cost of full GPT-5.4.<\/p>\n<p>If you&#8217;re routing by complexity \u2014 GPT-5.4 Mini handles the high-volume, low-latency tier while GPT-5.4 takes the reasoning-heavy work.<\/p>\n<div class=\"d-flex\"><a class=\"cta_button_link btn-secondary\" href=\"https:\/\/techcommunity.microsoft.com\/blog\/azure-ai-foundry-blog\/introducing-open-ai%E2%80%99s-gpt%E2%80%915-4-mini-in-microsoft-foundry\/4500569\" target=\"_blank\" rel=\"noopener\">Read the GPT-5.4 Mini Announcement<\/a><\/div>\n<h3>Phi-4 Reasoning Vision 15B<\/h3>\n<p><strong>Phi-4 Reasoning Vision 15B<\/strong> brings multimodal reasoning to the Phi family \u2014 a 15B-parameter model that combines visual understanding with chain-of-thought reasoning. Handles charts, diagrams, document layouts, and visual Q&amp;A with strong performance relative to its size.<\/p>\n<div class=\"d-flex\"><a class=\"cta_button_link btn-secondary\" href=\"https:\/\/aka.ms\/Phi-4-reasoning-vision-15B\" target=\"_blank\" rel=\"noopener\">Explore Phi-4 Reasoning Vision<\/a><\/div>\n<h3>Grok 4.2 (GA)<\/h3>\n<p><strong>Grok 4.2<\/strong> from xAI graduated to general availability on March 30, following its public beta earlier this year. A refreshed chat model in the catalog \u2014 available via serverless or provisioned throughput deployments.<\/p>\n<div class=\"d-flex\"><a class=\"cta_button_link btn-secondary\" href=\"https:\/\/ai.azure.com\/catalog\" target=\"_blank\" rel=\"noopener\">Explore in Model Catalog<\/a><\/div>\n<h3>Fireworks AI on Foundry (Public Preview)<\/h3>\n<p>Fireworks AI brings high-performance open model inference to Foundry \u2014 processing over 13 trillion tokens daily at ~180K requests\/second in production. Four models available at launch:<\/p>\n<table>\n<thead>\n<tr>\n<th>Model<\/th>\n<th>Notes<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><strong>DeepSeek V3.2<\/strong><\/td>\n<td>Sparse attention, 128K context<\/td>\n<\/tr>\n<tr>\n<td><strong>gpt-oss-120b<\/strong><\/td>\n<td>OpenAI&#8217;s open-source model<\/td>\n<\/tr>\n<tr>\n<td><strong>Kimi K2.5<\/strong><\/td>\n<td>Moonshot AI&#8217;s latest<\/td>\n<\/tr>\n<tr>\n<td><strong>MiniMax M2.5<\/strong><\/td>\n<td>New to Foundry with serverless support<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>The real story here is <strong>bring-your-own-weights (BYOW)<\/strong> \u2014 upload and register quantized or fine-tuned weights from anywhere without changing the serving stack. Deploy via serverless pay-per-token or provisioned throughput.<\/p>\n<div class=\"d-flex\"><a class=\"cta_button_link btn-secondary\" href=\"https:\/\/azure.microsoft.com\/en-us\/blog\/introducing-fireworks-ai-on-microsoft-foundry-bringing-high-performance-low-latency-open-model-inference-to-azure\/\" target=\"_blank\" rel=\"noopener\">Get Started with Fireworks<\/a><\/div>\n<h3>NVIDIA Nemotron Models<\/h3>\n<p>Announced at NVIDIA GTC on March 16, <strong>NVIDIA Nemotron models<\/strong> are now available through the Foundry catalog. Open models on NVIDIA accelerators, joining the widest selection of models on any cloud. Combined with Fireworks AI integration, you can fine-tune Nemotron into low-latency assets distributable to the edge.<\/p>\n<div class=\"d-flex\"><a class=\"cta_button_link btn-secondary\" href=\"https:\/\/ai.azure.com\/catalog\" target=\"_blank\" rel=\"noopener\">Browse Nemotron in Catalog<\/a><\/div>\n<h3>OSS Models in NextGen (GA)<\/h3>\n<p>Open-source models are now fully integrated into the NextGen Foundry experience at GA \u2014 unifying OSS and OpenAI models in a single deployment and management flow. No more context-switching between model providers.<\/p>\n<div class=\"d-flex\"><a class=\"cta_button_link btn-secondary\" href=\"https:\/\/ai.azure.com\/catalog\" target=\"_blank\" rel=\"noopener\">Explore Model Catalog<\/a><\/div>\n<hr \/>\n<h2>Agents<\/h2>\n<h3>Foundry Agent Service (GA)<\/h3>\n<p>The biggest story of the month. The <strong>next-gen Foundry Agent Service<\/strong> is generally available \u2014 built on the OpenAI Responses API, wire-compatible with OpenAI agents, and open to models from DeepSeek, xAI, Meta, LangChain, LangGraph, and more.<\/p>\n<p>If you&#8217;re building with the Responses API today, migrating to Foundry is minimal code changes. What you gain: enterprise security, private networking, Entra RBAC, full tracing, and evaluation \u2014 on top of your existing agent logic.<\/p>\n<pre><code class=\"language-python\">import os\r\nfrom azure.identity import DefaultAzureCredential\r\nfrom azure.ai.projects import AIProjectClient\r\nfrom azure.ai.projects.models import PromptAgentDefinition\r\n\r\nwith (\r\n    DefaultAzureCredential() as credential,\r\n    AIProjectClient(endpoint=os.environ[\"AZURE_AI_PROJECT_ENDPOINT\"], credential=credential) as project_client,\r\n    project_client.get_openai_client() as openai_client,\r\n):\r\n    agent = project_client.agents.create_version(\r\n        agent_name=\"my-enterprise-agent\",\r\n        definition=PromptAgentDefinition(\r\n            model=os.environ[\"AZURE_AI_MODEL_DEPLOYMENT_NAME\"],\r\n            instructions=\"You are a helpful assistant.\",\r\n        ),\r\n    )\r\n\r\n    conversation = openai_client.conversations.create()\r\n    response = openai_client.responses.create(\r\n        conversation=conversation.id,\r\n        input=\"What are best practices for building AI agents?\",\r\n        extra_body={\"agent_reference\": {\"name\": agent.name, \"type\": \"agent_reference\"}},\r\n    )\r\n    print(response.output_text)<\/code><\/pre>\n<blockquote><p><strong>Action:<\/strong> <code>pip install azure-ai-projects<\/code> \u2014 that&#8217;s it. As of 2.0.0, the package bundles <code>openai<\/code> and <code>azure-identity<\/code> as direct dependencies, so you no longer need to install them separately. If you&#8217;re coming from <code>azure-ai-agents<\/code>, agents are now first-class operations on <code>AIProjectClient<\/code> \u2014 remove your standalone <code>azure-ai-agents<\/code> pin.<\/p><\/blockquote>\n<div class=\"d-flex\"><a class=\"cta_button_link btn-secondary\" href=\"https:\/\/devblogs.microsoft.com\/foundry\/foundry-agent-service-ga\/\" target=\"_blank\" rel=\"noopener\">Read the Full Announcement<\/a><\/div>\n<h3>Voice Live + Foundry Agents (Preview)<\/h3>\n<p><strong>Voice Live<\/strong> is a fully managed, real-time speech-to-speech runtime that collapses the traditional STT \u2192 LLM \u2192 TTS pipeline into a single managed API. Semantic voice activity detection, end-of-turn detection, server-side noise suppression, echo cancellation, and barge-in support \u2014 all built-in.<\/p>\n<p>Connect Voice Live directly to an existing Foundry agent. The agent&#8217;s prompt, tools, and configuration stay in Foundry; Voice Live handles the audio pipeline. Voice interactions go through the same agent runtime as text \u2014 same evaluators, same traces, same cost visibility.<\/p>\n<pre><code class=\"language-python\">import asyncio\r\nfrom azure.ai.voicelive.aio import connect, AgentSessionConfig\r\nfrom azure.identity.aio import DefaultAzureCredential\r\n\r\nasync def run():\r\n    agent_config: AgentSessionConfig = {\r\n        \"agent_name\": \"my-enterprise-agent\",\r\n        \"project_name\": \"my-foundry-project\",\r\n    }\r\n\r\n    async with DefaultAzureCredential() as credential:\r\n        async with connect(\r\n            endpoint=os.environ[\"AZURE_VOICELIVE_ENDPOINT\"],\r\n            credential=credential,\r\n            agent_config=agent_config,\r\n        ) as connection:\r\n            await connection.session.update(session=session_config)\r\n            async for event in connection:\r\n                ...\r\n\r\nasyncio.run(run())<\/code><\/pre>\n<div class=\"d-flex\"><a class=\"cta_button_link btn-secondary\" href=\"https:\/\/github.com\/microsoft-foundry\/voicelive-samples\/blob\/main\/python\/voice-live-quickstarts\/AgentsNewQuickstart\/voice-live-with-agent-v2.py\" target=\"_blank\" rel=\"noopener\">View Voice Live Sample<\/a><\/div>\n<h3>MCP Authentication Expansion<\/h3>\n<p>MCP server connections now support the full authentication spectrum:<\/p>\n<table>\n<thead>\n<tr>\n<th>Auth Method<\/th>\n<th>Use Case<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><strong>Key-based<\/strong><\/td>\n<td>Static API tokens via Custom Keys connection<\/td>\n<\/tr>\n<tr>\n<td><strong>Entra Agent Identity<\/strong><\/td>\n<td>Service-to-service with managed identity<\/td>\n<\/tr>\n<tr>\n<td><strong>Managed Identity<\/strong><\/td>\n<td>Azure resource access<\/td>\n<\/tr>\n<tr>\n<td><strong>OAuth Identity Passthrough<\/strong><\/td>\n<td>User-delegated access (OneDrive, Salesforce, SaaS APIs)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>OAuth Identity Passthrough is the standout \u2014 when agents need to act on behalf of a specific user, not as a shared system identity:<\/p>\n<pre><code class=\"language-python\">from azure.ai.projects.models import MCPTool, PromptAgentDefinition\r\n\r\ntool = MCPTool(\r\n    server_label=\"github-api\",\r\n    server_url=\"https:\/\/api.githubcopilot.com\/mcp\",\r\n    require_approval=\"always\",\r\n    project_connection_id=os.environ[\"MCP_PROJECT_CONNECTION_ID\"],\r\n)\r\n\r\nagent = project_client.agents.create_version(\r\n    agent_name=\"my-mcp-agent\",\r\n    definition=PromptAgentDefinition(\r\n        model=os.environ[\"AZURE_AI_MODEL_DEPLOYMENT_NAME\"],\r\n        instructions=\"Use MCP tools as needed.\",\r\n        tools=[tool],\r\n    ),\r\n)<\/code><\/pre>\n<div class=\"d-flex\"><a class=\"cta_button_link btn-secondary\" href=\"https:\/\/learn.microsoft.com\/en-us\/azure\/ai-foundry\/agents\/how-to\/tools\/model-context-protocol\" target=\"_blank\" rel=\"noopener\">MCP Auth Documentation<\/a><\/div>\n<h3>Evaluations (GA) + Continuous Monitoring<\/h3>\n<p>Foundry Evaluations shipped GA with three layers:<\/p>\n<ul>\n<li><strong>Out-of-the-box evaluators<\/strong> \u2014 coherence, relevance, groundedness, retrieval quality, safety. No configuration required.<\/li>\n<li><strong>Custom evaluators<\/strong> \u2014 encode your own criteria: business logic, tone standards, domain compliance.<\/li>\n<li><strong>Continuous evaluation<\/strong> \u2014 Foundry samples live traffic automatically, runs your evaluator suite, and surfaces results in Azure Monitor dashboards. Configure alerts for quality drift.<\/li>\n<\/ul>\n<p>Here&#8217;s the pattern for running evaluations against an agent target:<\/p>\n<pre><code class=\"language-python\">eval_object = openai_client.evals.create(\r\n    name=\"Agent Quality Evaluation\",\r\n    data_source_config=DataSourceConfigCustom(\r\n        type=\"custom\",\r\n        item_schema={\"type\": \"object\", \"properties\": {\"query\": {\"type\": \"string\"}}, \"required\": [\"query\"]},\r\n        include_sample_schema=True,\r\n    ),\r\n    testing_criteria=[\r\n        {\r\n            \"type\": \"azure_ai_evaluator\",\r\n            \"name\": \"fluency\",\r\n            \"evaluator_name\": \"builtin.fluency\",\r\n            \"initialization_parameters\": {\"deployment_name\": os.environ[\"AZURE_AI_MODEL_DEPLOYMENT_NAME\"]},\r\n            \"data_mapping\": {\"query\": \"{{item.query}}\", \"response\": \"{{sample.output_text}}\"},\r\n        },\r\n        {\r\n            \"type\": \"azure_ai_evaluator\",\r\n            \"name\": \"task_adherence\",\r\n            \"evaluator_name\": \"builtin.task_adherence\",\r\n            \"initialization_parameters\": {\"deployment_name\": os.environ[\"AZURE_AI_MODEL_DEPLOYMENT_NAME\"]},\r\n            \"data_mapping\": {\"query\": \"{{item.query}}\", \"response\": \"{{sample.output_items}}\"},\r\n        },\r\n    ],\r\n)\r\n\r\nrun = openai_client.evals.runs.create(\r\n    eval_id=eval_object.id,\r\n    name=f\"Run for {agent.name}\",\r\n    data_source={\r\n        \"type\": \"azure_ai_target_completions\",\r\n        \"source\": {\r\n            \"type\": \"file_content\",\r\n            \"content\": [{\"item\": {\"query\": \"What is the capital of France?\"}}],\r\n        },\r\n        \"input_messages\": {\r\n            \"type\": \"template\",\r\n            \"template\": [{\"type\": \"message\", \"role\": \"user\", \"content\": {\"type\": \"input_text\", \"text\": \"{{item.query}}\"}}],\r\n        },\r\n        \"target\": {\"type\": \"azure_ai_agent\", \"name\": agent.name, \"version\": agent.version},\r\n    },\r\n)<\/code><\/pre>\n<div class=\"d-flex\"><a class=\"cta_button_link btn-secondary\" href=\"https:\/\/github.com\/Azure\/azure-sdk-for-python\/tree\/main\/sdk\/ai\/azure-ai-projects\/samples\/evaluations\" target=\"_blank\" rel=\"noopener\">View Evaluation Samples<\/a><\/div>\n<h3>Prompt Optimizer in Agent Playground<\/h3>\n<p>The <strong>Prompt Optimizer<\/strong> is now integrated directly into the Agent playground \u2014 iteratively improve prompts before shipping agents. Data-driven prompt tuning connected to evaluation results.<\/p>\n<div class=\"d-flex\"><a class=\"cta_button_link btn-secondary\" href=\"https:\/\/ai.azure.com\" target=\"_blank\" rel=\"noopener\">Try Prompt Optimizer<\/a><\/div>\n<h3>Hosted Agents in 6 New Regions<\/h3>\n<p>Hosted agent deployments are now available in <strong>East US, North Central US, Sweden Central, Southeast Asia, Japan East, and more<\/strong> \u2014 relevant for data residency requirements and latency optimization.<\/p>\n<div class=\"d-flex\"><a class=\"cta_button_link btn-secondary\" href=\"https:\/\/learn.microsoft.com\/azure\/foundry\/agents\/quickstarts\/quickstart-hosted-agent\" target=\"_blank\" rel=\"noopener\">Deploy a Hosted Agent<\/a><\/div>\n<hr \/>\n<h2>Safety &amp; Guardrails<\/h2>\n<h3>Palo Alto Prisma AIRS + Zenity (GA)<\/h3>\n<p>Third-party guardrail integrations are now generally available. Register and manage runtime security from <strong>Palo Alto Networks Prisma AIRS<\/strong> and <strong>Zenity<\/strong> directly in Foundry to detect prompt injection, toxic content, malicious URLs, sensitive data leakage, and tool misuse alongside native guardrails.<\/p>\n<table>\n<thead>\n<tr>\n<th>Integration<\/th>\n<th>Detects<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><strong>Palo Alto Prisma AIRS<\/strong><\/td>\n<td>Prompt injection, toxic content, malicious URLs, data leakage<\/td>\n<\/tr>\n<tr>\n<td><strong>Zenity<\/strong><\/td>\n<td>Prompt injection, tool misuse, data exfiltration<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>Announced at NVIDIA GTC alongside the Foundry Agent Service GA.<\/p>\n<div class=\"d-flex\"><a class=\"cta_button_link btn-secondary\" href=\"https:\/\/learn.microsoft.com\/azure\/foundry\/guardrails\/guardrails-overview\" target=\"_blank\" rel=\"noopener\">Guardrails Documentation<\/a><\/div>\n<h3>Task Adherence as Native Guardrail<\/h3>\n<p><strong>Task Adherence<\/strong> is now a native guardrail risk type in Foundry \u2014 block or annotate off-task tool calls without wiring directly to the standalone Task Adherence API.<\/p>\n<div class=\"d-flex\"><a class=\"cta_button_link btn-secondary\" href=\"https:\/\/learn.microsoft.com\/azure\/foundry\/guardrails\/guardrails-overview\" target=\"_blank\" rel=\"noopener\">Configure Task Adherence<\/a><\/div>\n<h3>Tool-Call &amp; Tool-Response Guardrails<\/h3>\n<p>New intervention points in Foundry Guardrails for <strong>tool invocations and tool responses<\/strong> \u2014 detect and mitigate risks in what tools are being called and what they return. Reduces data leakage and unsafe tool behavior in agentic workflows.<\/p>\n<div class=\"d-flex\"><a class=\"cta_button_link btn-secondary\" href=\"https:\/\/learn.microsoft.com\/azure\/foundry\/guardrails\/guardrails-overview\" target=\"_blank\" rel=\"noopener\">Explore Tool Guardrails<\/a><\/div>\n<h3>Agent Mitigations &amp; Guardrail Customization (GA)<\/h3>\n<p>Hardened agent mitigations and guardrail customization round out GA-level safety controls for production agents.<\/p>\n<div class=\"d-flex\"><a class=\"cta_button_link btn-secondary\" href=\"https:\/\/learn.microsoft.com\/azure\/foundry\/guardrails\/guardrails-overview\" target=\"_blank\" rel=\"noopener\">Safety Controls Documentation<\/a><\/div>\n<hr \/>\n<h2>Speech, Audio &amp; Avatars<\/h2>\n<h3>Neural HD TTS Updates + MAI-voice-1<\/h3>\n<p>March brings updates to the <strong>Neural HD TTS stack<\/strong> including MAI-voice-1 integration \u2014 higher-fidelity, more expressive synthetic speech with improved quality across secondary locales.<\/p>\n<div class=\"d-flex\"><a class=\"cta_button_link btn-secondary\" href=\"https:\/\/learn.microsoft.com\/en-us\/azure\/ai-services\/speech-service\/text-to-speech\" target=\"_blank\" rel=\"noopener\">TTS Documentation<\/a><\/div>\n<h3>Fast Transcription \u2014 5-Hour Support<\/h3>\n<p><strong>Fast Transcription<\/strong> now supports up to ~5-hour audio inputs \u2014 addressing top customer requests for long-form meeting and media transcription.<\/p>\n<div class=\"d-flex\"><a class=\"cta_button_link btn-secondary\" href=\"https:\/\/learn.microsoft.com\/en-us\/azure\/ai-services\/speech-service\/batch-transcription\" target=\"_blank\" rel=\"noopener\">Transcription Documentation<\/a><\/div>\n<h3>Dynamic Vocabulary (GA)<\/h3>\n<p><strong>Dynamic vocabulary for English<\/strong> is now GA \u2014 improving recognition of domain-specific terms in Teams transcription and related STT workloads. Custom dictionary support (Tier 2) is in public preview for finer pronunciation control.<\/p>\n<div class=\"d-flex\"><a class=\"cta_button_link btn-secondary\" href=\"https:\/\/learn.microsoft.com\/en-us\/azure\/ai-services\/speech-service\/speech-to-text\" target=\"_blank\" rel=\"noopener\">STT Documentation<\/a><\/div>\n<h3>Custom Photo &amp; Video Avatars in NextGen<\/h3>\n<p>Bring <strong>custom photo and video avatars<\/strong> into the Foundry NextGen experience \u2014 supporting richer, branded AI presenters for enterprise scenarios.<\/p>\n<div class=\"d-flex\"><a class=\"cta_button_link btn-secondary\" href=\"https:\/\/learn.microsoft.com\/en-us\/azure\/ai-services\/speech-service\/text-to-speech-avatar\/what-is-text-to-speech-avatar\" target=\"_blank\" rel=\"noopener\">Avatar Documentation<\/a><\/div>\n<h3>Playground GAs \u2014 TTS, Avatar, STT &amp; Speech Translation<\/h3>\n<p>Three playgrounds reached GA this month:<\/p>\n<table>\n<thead>\n<tr>\n<th>Playground<\/th>\n<th>Status<\/th>\n<th>What It Does<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><strong>TTS Playground<\/strong><\/td>\n<td>GA<\/td>\n<td>Audition voices and parameters before integrating<\/td>\n<\/tr>\n<tr>\n<td><strong>Avatar Playground<\/strong><\/td>\n<td>GA<\/td>\n<td>Preview and tune avatar experiences in NextGen<\/td>\n<\/tr>\n<tr>\n<td><strong>STT &amp; Speech Translation Playground<\/strong><\/td>\n<td>GA<\/td>\n<td>Trial STT and translation models<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<div class=\"d-flex\"><a class=\"cta_button_link btn-secondary\" href=\"https:\/\/ai.azure.com\" target=\"_blank\" rel=\"noopener\">Try the Playgrounds<\/a><\/div>\n<hr \/>\n<h2>Platform<\/h2>\n<h3>Priority Processing (Preview)<\/h3>\n<p><strong>Priority Processing<\/strong> gives you a dedicated compute lane for latency-sensitive and business-critical AI workloads. When you need guaranteed low-latency inference \u2014 real-time agents, customer-facing chat, time-sensitive pipelines \u2014 Priority Processing routes your requests through reserved capacity instead of competing with standard traffic.<\/p>\n<p>Available for OpenAI models in Foundry with configurable priority tiers per deployment.<\/p>\n<div class=\"d-flex\"><a class=\"cta_button_link btn-secondary\" href=\"https:\/\/techcommunity.microsoft.com\/blog\/azure-ai-foundry-blog\/announcing-priority-processing-in-microsoft-foundry-for-performance-sensitive-ai\/4504788\" target=\"_blank\" rel=\"noopener\">Read the Priority Processing Announcement<\/a><\/div>\n<h3>Tracing (GA)<\/h3>\n<p>Tracing is finalized for GA \u2014 UX polish, data model refinements, and sort\/filter capabilities so you can reliably inspect agent traces in production. New OTel semantics for AI workloads (memory, state, planning) improve interoperability across tooling.<\/p>\n<div class=\"d-flex\"><a class=\"cta_button_link btn-secondary\" href=\"https:\/\/learn.microsoft.com\/en-us\/azure\/ai-foundry\/how-to\/develop\/visualize-traces\" target=\"_blank\" rel=\"noopener\">Tracing Documentation<\/a><\/div>\n<h3>End-to-End Private Networking<\/h3>\n<p>Foundry Agent Service now supports <strong>Standard Setup with private networking<\/strong> \u2014 BYO VNet, no public egress, container\/subnet injection. Extended to tool connectivity: MCP servers, Azure AI Search indexes, and Fabric data agents all operate over private network paths.<\/p>\n<p>Managed VNET logging adds firewall\/NSG\/flow logs for visibility into isolated environments.<\/p>\n<div class=\"d-flex\"><a class=\"cta_button_link btn-secondary\" href=\"https:\/\/learn.microsoft.com\/en-us\/azure\/ai-foundry\/agents\/how-to\/virtual-networks\" target=\"_blank\" rel=\"noopener\">Private Networking Docs<\/a><\/div>\n<h3>Foundry Control Plane ARM API<\/h3>\n<p>A consolidated <strong>Foundry Control Plane ARM API<\/strong> gives enterprises a unified way to manage agents, models, and tools via ARM. Public preview of FCP support for <strong>Azure Functions and App Service<\/strong> lets you govern function-hosted agents centrally.<\/p>\n<div class=\"d-flex\"><a class=\"cta_button_link btn-secondary\" href=\"https:\/\/learn.microsoft.com\/azure\/foundry\/control-plane\/how-to-manage-agents\" target=\"_blank\" rel=\"noopener\">FCP Documentation<\/a><\/div>\n<h3>Fine-Tuning CLI<\/h3>\n<p>A new <strong>CLI for configuring, submitting, and monitoring fine-tuning jobs<\/strong> \u2014 a faster, code-first way to iterate on custom models. Cost estimation based on token projections helps plan spend before running large training jobs.<\/p>\n<div class=\"d-flex\"><a class=\"cta_button_link btn-secondary\" href=\"https:\/\/learn.microsoft.com\/en-us\/azure\/ai-foundry\/concepts\/fine-tuning-overview\" target=\"_blank\" rel=\"noopener\">Fine-Tuning Documentation<\/a><\/div>\n<h3>Platform Updates Rollup<\/h3>\n<ul>\n<li><strong>Eval results \u2194 agent trace linking<\/strong> \u2014 evaluation results now connect to the underlying agent trace, closing a key observability gap for debugging.<\/li>\n<li><strong>Local evals aligned with Evaluators catalog<\/strong> \u2014 run local evaluations with the same primitives as hosted runs, no hardcoded SDK logic.<\/li>\n<li><strong>PII NextGen playgrounds<\/strong> \u2014 conversational and document PII detection with updated configuration panels exposing preview features.<\/li>\n<li><strong>Notification center<\/strong> \u2014 tenant-level notifications (not just project-scoped), plus email delivery for critical eval, safety, and deployment alerts.<\/li>\n<li><strong>Free Trial &amp; PAYG<\/strong> \u2014 Free Trial as default sign-up path, in-app PAYG subscription creation, and in-app trial start to reduce friction.<\/li>\n<li><strong>CMK for Azure AI Search<\/strong> \u2014 service-level customer-managed key configuration so admins set encryption defaults once, not per-index.<\/li>\n<\/ul>\n<hr \/>\n<h2>SDK &amp; Language Changelog (March 2026)<\/h2>\n<p>March was the SDK GA month. The Foundry REST API went GA in February, and this month the SDKs followed \u2014 <strong>Python, JS\/TS, and Java all shipped stable 2.0.0 releases<\/strong> targeting the v1 REST surface. .NET 2.0.0 shipped April 1. The <code>azure-ai-agents<\/code> dependency is gone across all languages; agents, evals, memory, and inference all live under the unified <code>AIProjectClient<\/code>.<\/p>\n<h3>Python<\/h3>\n<p><strong><code>azure-ai-projects<\/code> 2.0.0 (Mar 6) + 2.0.1 (Mar 12)<\/strong><\/p>\n<p>First stable release. This is the one to pin for production.<\/p>\n<p><strong>Features:<\/strong><\/p>\n<ul>\n<li><strong>Dependency consolidation:<\/strong> <code>azure-ai-projects<\/code> now bundles <code>openai<\/code> and <code>azure-identity<\/code> as direct dependencies \u2014 <code>pip install azure-ai-projects<\/code> is the only install command you need. No more juggling three packages.<\/li>\n<li>New <code>allow_preview<\/code> boolean on <code>AIProjectClient<\/code> constructor replaces per-method <code>foundry_features<\/code> \u2014 opt in once for all preview operations<\/li>\n<li>Preview operations (hosted agents, workflow agents) use the same <code>allow_preview<\/code> flag; <code>.beta<\/code> sub-client methods imply it automatically<\/li>\n<\/ul>\n<p><strong>Breaking changes from 2.0.0b4:<\/strong><\/p>\n<pre><code class=\"language-python\"># Before \u2014 per-method foundry_features\r\nagent = project_client.agents.create_version(\r\n    model=\"gpt-5\",\r\n    foundry_features=FoundryFeaturesOptInKeys.WORKFLOW_AGENTS_V1_PREVIEW,\r\n)\r\n\r\n# After \u2014 constructor-level allow_preview\r\nproject_client = AIProjectClient(\r\n    endpoint=os.environ[\"AZURE_AI_PROJECT_ENDPOINT\"],\r\n    credential=credential,\r\n    allow_preview=True,  # enables all preview features\r\n)\r\nagent = project_client.agents.create_version(model=\"gpt-5\")<\/code><\/pre>\n<p>Other renames:<\/p>\n<ul>\n<li><code>TextResponseFormatConfiguration<\/code> \u2192 <code>TextResponseFormat<\/code><\/li>\n<li><code>CodeInterpreterContainerAuto<\/code> \u2192 <code>AutoCodeInterpreterToolParam<\/code> (+ new <code>network_policy<\/code> property)<\/li>\n<li><code>ImageGenActionEnum<\/code> \u2192 <code>ImageGenAction<\/code><\/li>\n<li>Datetime fields across <code>CronTrigger<\/code>, <code>RecurrenceTrigger<\/code>, <code>OneTimeTrigger<\/code>, <code>ScheduleRun<\/code> changed from <code>str<\/code> to <code>datetime.datetime<\/code><\/li>\n<\/ul>\n<blockquote><p><strong>Action:<\/strong> <code>pip install azure-ai-projects==2.0.1<\/code> \u2014 pin to the stable release. If you were on 2.0.0b4, replace <code>foundry_features<\/code> with <code>allow_preview=True<\/code> on the client constructor.<\/p><\/blockquote>\n<p><a href=\"https:\/\/pypi.org\/project\/azure-ai-projects\/2.0.1\/\">Changelog<\/a><\/p>\n<h3>.NET<\/h3>\n<p><strong><code>Azure.AI.Projects<\/code> 2.0.0-beta.2 (Mar 12)<\/strong><\/p>\n<p>The .NET SDK restructured packages \u2014 agents administration moved to <code>Azure.AI.Projects.Agents<\/code>, and <code>Azure.AI.Projects.OpenAI<\/code> was renamed to <code>Azure.AI.Extensions.OpenAI<\/code>. OpenAI dependency upgraded to 2.9.1.<\/p>\n<blockquote><p><code>.NET 2.0.0 GA shipped April 1<\/code> \u2014 the first .NET stable release on the v1 REST surface. Major renames: <code>Insights<\/code> \u2192 <code>ProjectInsights<\/code>, evaluations\/memory moved to separate namespaces, <code>AIProjectClient.OpenAI<\/code> \u2192 <code>AIProjectClient.ProjectOpenAIClient<\/code>, <code>AIProjectClient.Agents<\/code> \u2192 <code>AIProjectClient.AgentAdministrationClient<\/code>.<\/p>\n<p><strong>Action:<\/strong> Upgrade to <code>Azure.AI.Projects 2.0.0<\/code> (GA, April 1). Review the <a href=\"https:\/\/github.com\/Azure\/azure-sdk-for-net\/blob\/main\/sdk\/ai\/Azure.AI.Projects\/CHANGELOG.md\">breaking changes<\/a> \u2014 significant property and namespace renames.<\/p><\/blockquote>\n<p><a href=\"https:\/\/github.com\/Azure\/azure-sdk-for-net\/blob\/main\/sdk\/ai\/Azure.AI.Projects\/CHANGELOG.md\">Changelog<\/a><\/p>\n<h3>JavaScript \/ TypeScript<\/h3>\n<p><strong><code>@azure\/ai-projects<\/code> 2.0.0 (Mar 6) + 2.0.1 (Mar 13)<\/strong><\/p>\n<p>First stable release for JS\/TS.<\/p>\n<p><strong>Breaking changes from 2.0.0-beta.5:<\/strong><\/p>\n<ul>\n<li><code>RedTeam.target<\/code> changed from required to optional<\/li>\n<li><code>container_app<\/code> removed from <code>AgentKind<\/code>; <code>ContainerAppAgentDefinition<\/code> removed<\/li>\n<li><code>project.connections.get<\/code> and <code>.getDefault<\/code> \u2014 <code>includeCredentials<\/code> moved to options bag<\/li>\n<li><code>project.beta.evaluators.listLatestVersions<\/code> \u2192 <code>project.beta.evaluators.list<\/code><\/li>\n<\/ul>\n<blockquote><p><strong>Action:<\/strong> <code>npm install @azure\/ai-projects@2.0.1<\/code> \u2014 pin to stable. The beta \u2192 GA migration is mostly renames and options bag changes.<\/p><\/blockquote>\n<p><a href=\"https:\/\/www.npmjs.com\/package\/@azure\/ai-projects\">Changelog<\/a><\/p>\n<h3>Java<\/h3>\n<p><strong><code>azure-ai-projects<\/code> 2.0.0-beta.2 (Mar 4) \u2192 2.0.0-beta.3 (Mar 19) \u2192 2.0.0 (Mar 27)<\/strong><\/p>\n<p>Three releases in March, culminating in the first Java GA. The beta releases iterated on breaking changes before locking the API surface.<\/p>\n<p><strong>Key breaking changes in 2.0.0:<\/strong><\/p>\n<ul>\n<li>Method renames across all sub-clients for disambiguation (e.g., <code>list()<\/code> \u2192 <code>listDeployments()<\/code>, <code>get()<\/code> \u2192 <code>getDeployment()<\/code>)<\/li>\n<li><code>Connection.getCredentials()<\/code> \u2192 <code>Connection.getCredential()<\/code> (singular)<\/li>\n<li><code>FoundryFeaturesOptInKeys<\/code> changed from <code>ExpandableStringEnum<\/code> to standard Java <code>enum<\/code><\/li>\n<li><code>DatasetsClient.createDatasetWithFolder()<\/code> throws <code>UncheckedIOException<\/code> instead of checked <code>IOException<\/code><\/li>\n<li><code>DatasetVersion.getDataUri()<\/code> \u2192 <code>getDataUrl()<\/code><\/li>\n<\/ul>\n<blockquote><p><strong>Action:<\/strong> <code>mvn dependency:resolve -Dartifact=com.azure:azure-ai-projects:2.0.0<\/code> \u2014 pin to stable. Review the full changelog for the method rename table.<\/p><\/blockquote>\n<p><a href=\"https:\/\/github.com\/Azure\/azure-sdk-for-java\/blob\/main\/sdk\/ai\/azure-ai-projects\/CHANGELOG.md\">Changelog<\/a><\/p>\n<hr \/>\n<h2>Deprecations<\/h2>\n<p>Plan your migrations now \u2014 these timelines are firm.<\/p>\n<table>\n<thead>\n<tr>\n<th>Deprecation<\/th>\n<th>Migration Target<\/th>\n<th>Deadline<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><strong>PromptFlow<\/strong> (Azure AI Foundry + Azure ML)<\/td>\n<td>Microsoft Framework Workflows<\/td>\n<td>January 2027<\/td>\n<\/tr>\n<tr>\n<td><strong>Import Data \/ Data Connections<\/strong> (Azure ML)<\/td>\n<td>Fabric OneLake patterns<\/td>\n<td>Effective now<\/td>\n<\/tr>\n<tr>\n<td><strong>Low-priority VMs<\/strong> (Azure ML)<\/td>\n<td>Spot VMs<\/td>\n<td>Effective now<\/td>\n<\/tr>\n<tr>\n<td><strong>Default internet access<\/strong> for new managed VNets<\/td>\n<td>Explicit outbound configuration<\/td>\n<td>Effective March 31, 2026<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<blockquote><p><strong>Action:<\/strong> If you&#8217;re using PromptFlow in production, start planning your migration to <a href=\"https:\/\/learn.microsoft.com\/en-us\/agent-framework\">Microsoft Framework Workflows<\/a>. The January 2027 sunset gives you nine months.<\/p><\/blockquote>\n<div class=\"d-flex\"><a class=\"cta_button_link btn-secondary\" href=\"https:\/\/learn.microsoft.com\/en-us\/agent-framework\" target=\"_blank\" rel=\"noopener\">Explore Microsoft Framework Workflows<\/a><\/div>\n<hr \/>\n<h2>Resources &amp; Community<\/h2>\n<p><div class=\"alert alert-info\"><p class=\"alert-divider\"><i class=\"fabric-icon fabric-icon--Info\"><\/i><strong>Forrester TEI Study: The Economics of Enterprise AI<\/strong><\/p>A new Forrester Total Economic Impact study found that organizations using Microsoft Foundry saw <strong>20\u201330% developer time savings<\/strong> and a <strong>sub-6-month payback period<\/strong>. If you&#8217;re building the business case for standardizing on Foundry, these are the numbers. <a href=\"https:\/\/azure.microsoft.com\/en-us\/blog\/the-economics-of-enterprise-ai-what-the-forrester-tei-study-reveals-about-microsoft-foundry\/\">Read the full study \u2192<\/a><\/div><\/p>\n<ul>\n<li><strong>Discord:<\/strong> Join 50,000+ developers in <a href=\"https:\/\/aka.ms\/foundry\/discord\">the Foundry Discord<\/a><\/li>\n<li><strong>GitHub Discussions:<\/strong> Ask questions in <a href=\"https:\/\/aka.ms\/foundry\/forum\">the forum<\/a><\/li>\n<li><strong>RSS:<\/strong> <a href=\"https:\/\/devblogs.microsoft.com\/foundry\/category\/whats-new\/feed\/\">Subscribe<\/a> to get this digest monthly<\/li>\n<li><strong>Model Mondays:<\/strong> <a href=\"https:\/\/aka.ms\/model-mondays\">Tune in live on YouTube<\/a> \u2014 Fireworks AI joined on March 23<\/li>\n<li><strong>The Shift podcast:<\/strong> <a href=\"https:\/\/www.youtube.com\/playlist?list=PLLasX02E8BPBCP7KdYsjKKFFQUmNEUmE9\">Listen and subscribe<\/a> for deep dives on agentic AI<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>\ufeffMarch ships Foundry Agent Service GA with private networking, GPT-5.4 and GPT-5.4 Mini, Priority Processing, Phi-4 Reasoning Vision, SDK 2.0 GA across Python, JS\/TS, Java, and .NET, Fireworks AI and NVIDIA Nemotron models, and third-party guardrails from Palo Alto and Zenity.<\/p>\n","protected":false},"author":185793,"featured_media":2101,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[1,27],"tags":[87,25,66,116,115,118,113,2,103,119,120,104,114],"class_list":["post-2084","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-microsoft-foundry","category-whats-new","tag-agent-framework","tag-agents","tag-evaluations","tag-fireworks","tag-gpt-5-4","tag-gpt-5-4-mini","tag-guardrails","tag-microsoft-foundry","tag-models","tag-phi-4","tag-priority-processing","tag-sdk","tag-speech"],"acf":[],"blog_post_summary":"<p>\ufeffMarch ships Foundry Agent Service GA with private networking, GPT-5.4 and GPT-5.4 Mini, Priority Processing, Phi-4 Reasoning Vision, SDK 2.0 GA across Python, JS\/TS, Java, and .NET, Fireworks AI and NVIDIA Nemotron models, and third-party guardrails from Palo Alto and Zenity.<\/p>\n","_links":{"self":[{"href":"https:\/\/devblogs.microsoft.com\/foundry\/wp-json\/wp\/v2\/posts\/2084","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/devblogs.microsoft.com\/foundry\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/devblogs.microsoft.com\/foundry\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/devblogs.microsoft.com\/foundry\/wp-json\/wp\/v2\/users\/185793"}],"replies":[{"embeddable":true,"href":"https:\/\/devblogs.microsoft.com\/foundry\/wp-json\/wp\/v2\/comments?post=2084"}],"version-history":[{"count":1,"href":"https:\/\/devblogs.microsoft.com\/foundry\/wp-json\/wp\/v2\/posts\/2084\/revisions"}],"predecessor-version":[{"id":2102,"href":"https:\/\/devblogs.microsoft.com\/foundry\/wp-json\/wp\/v2\/posts\/2084\/revisions\/2102"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/devblogs.microsoft.com\/foundry\/wp-json\/wp\/v2\/media\/2101"}],"wp:attachment":[{"href":"https:\/\/devblogs.microsoft.com\/foundry\/wp-json\/wp\/v2\/media?parent=2084"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/devblogs.microsoft.com\/foundry\/wp-json\/wp\/v2\/categories?post=2084"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/devblogs.microsoft.com\/foundry\/wp-json\/wp\/v2\/tags?post=2084"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}