Microsoft Foundry changelog

Claude costs appear under Foundry resources in Azure Cost Management, with tagging supportGA

Native triggers for published agents: recurring schedules, one-shot timers and event-based triggersPreview

Admins can enable or disable Foundry agents from Microsoft Admin CenterGA

Ordered allow, deny, transform and rewrite rules for hosted agents' outbound connections, with audit and enforce modesPreview

Publish Foundry agents directly to Microsoft 365 Copilot and Teams, governed through Entra and Agent 365GA

Lets developers and partners build and distribute custom azd extensionsGA

Sequential, parallel and Magentic orchestration patterns for .NET and Python, with SLA-backed supportGA

Production agent runtime for .NET and Python, with OpenTelemetry traces and SLA-backed supportGA

Multi-agent orchestration SDK for C# and Python, consolidating AutoGen and Semantic Kernel patterns

OpenTelemetry-compatible tracing for Python and .NET agents, routable to Application InsightsPreview

Agents run multi-step plans as one code block in Hyperlight sandboxed containersPreview

Agents store successful task patterns as episodes and replay them for similar requestsPreview

Local debugger showing agent runs as a tree of model calls and tool invocationsPreview

Playground in Foundry NextGen portal shows redacted output beside source for a built-in sample documentPreview

Refreshed playground with curated sample inputs and outputs for evaluating PII detection on documentsGA

Governed tool bundles exposed through a single MCP-compatible endpointGA

Custom voices can be paired with real-time avatars, removing the curated voice set limitPreview

Azure SQL Database added as a knowledge source in Azure AI SearchPreview

Azure API Management can front all Foundry model integrations in Azure AI SearchPreview

Retrieves only the most relevant tools at runtime from large toolbox catalogsPreview

Managed knowledge layer that grounds agents in enterprise data with ingestion, indexing and user-permission enforcementGA

Fabric Ontology can be queried as a federated knowledge source by agentsPreview

Private end-to-end connectivity to Foundry services via Shared Private Link or Network Security PerimeterGA

Adds semantic chunking and text descriptions of images, charts and diagrams during indexingPreview

First-party voice model with speech in 10+ languages, voice cloning and voice promptingPreview

Captures end user feedback on AI agents and links it to request traces in Azure MonitorPreview

NextGen playground in the Foundry portal extracts clinical entities with UMLS linking and FHIR-structured outputGA

Self-serve creation of custom photo avatars in the Foundry NextGen fine-tuning portalGA

Agent-to-agent communication for Prompt and Hosted agents using the responses protocolPreview

Weighted-dimension rubrics measure agent quality in single-turn and multi-turn flows, with optional human reviewPreview

Custom Voice authoring moves into Foundry portal, replacing the Speech Studio flowGA

Evaluations run on a diverse subset of production traces chosen by a MinHash farthest-first algorithmPreview

Extension adds model catalog of 1,900+ models, playground and Hosted agent deployment in the editorGA

Domain filter narrows 1,900+ models by use case such as robotics or coding, also via REST APIPreview

azd evaluation experience adds a quality loop to Foundry hosted agents, from first run to reusePreview

Updated playground with refreshed API Configuration Panel for testing PII categories, redaction modes and languagesPreview

New playground with transcript input and API Configuration Panel for testing PII detection in multi-speaker dialoguePreview

Observe skill in Foundry Plugin runs evaluation-driven agent optimization loop from the editorPreview

Unified API fronts neural, LLM, small language model, Adaptive custom and Custom Translator enginesGA

One reservation now covers Global PTU deployments across multiple regionsPrice

Connects real-time speech-to-text and text-to-speech to agents without a custom audio pipelineGA

Built-in policies govern deployment regions, approved models, routing rules and logging for model routerPreview

Run open-source benchmarks against model deployments or agents, compare runs in evaluation group viewPreview

Stable release across C#, C++, Java, JavaScript, Python, Objective-C, Swift and GoGA

Region-agnostic reservations let one PTU commitment be used in any supported regionGA

OneLake Catalog added as a knowledge source, respecting Fabric permissions and sensitivity labelsPreview

Knowledge bases return document-extracted images with retrieval results at query timePreview

Managed identity authentication for billing calls to Foundry resources, replacing key-based flowGA

Indexers call chat completion models during enrichment; chat completion also in knowledge sourcesGA

Setting chooses layout-aware or faster text-only document analysis, in API version 2025-11-01GA

Indexer adds ASPX site content and SharePoint Lists alongside document librariesPreview

Audit events for Microsoft Purview sensitivity labels flow into Purview activity logsPreview

Standard and custom talking head video from a single photo, standard at 512x512GA

Adapts translations to domain terminology using small datasets without model retraining, in NextGen playgroundGA

SDKs for Translator API v4 in C#/.NET, Java, JavaScript and PythonGA

Synchronous single-document image translation through one API call, preserving original layoutGA

Read and Layout playgrounds and prebuilt analyzers added to the Foundry portalPreview

Incremental sync keeps SharePoint document, site and library ACLs current without full re-indexingPreview

External MCP servers can be used as federated knowledge sources in agentic retrievalPreview

Step-by-step replay and new trace visualizations for debugging multi-step agent runsPreview

Authoring for custom video avatars and Photo Avatar added to the Foundry portalGA

Connectors and time-based triggers let agents react to events and schedulesPreview

Updated text-to-speech model replaces DragonHDLatest, with style tags for English voices and lower word error ratePreview

Serverless option for native indexers that scales execution with workload demandPreview

New Microsoft image generation model with better photorealism, prompt fidelity and speed than MAI-Image-2Preview

New speech-to-text model with lower word error rate on long-tail locales and faster transcriptionPreview

Purview sensitivity labels flow through ingestion into knowledge sources and knowledge basesPreview

Container-based agents run in microVM-isolated sandboxes with session starts under 100 millisecondsGA

Admin operations on Purview-labeled content flow into Purview's unified audit logPreview

Redesigned storage model, retrieval controls, semantic deduplication, TTL controls and content safety filtering on writePreview

LLM-powered transcription and translation for audio files across 25 languages, 90+ localesGA

Run cloud evaluation jobs against models and agents behind private endpointsGA

Register a OneLake item once and reuse it across knowledge sources and agentsGA

Batch document translation translates text inside images embedded in Word .docx filesGA

Batch document translation accepts .jpeg, .png, .bmp and .webp images, priced per imageGA

Better structure recovery for digital and scanned PDFs using Document IntelligenceGA

Azure AI roles renamed to Foundry roles; owner roles can assign select dependent roles

Version 1.0 for .NET and Python with stable APIs and long-term supportGA

Managed long-term memory for agents, integrates with Microsoft Agent Framework and LangGraphPreview

Agents can use models deployed outside Foundry, including behind gateways such as Azure API ManagementGA

Adds hosted-agent support to create, debug, evaluate and deploy agents in the editorGA

Runs agents in isolated sandboxes with a dedicated runtime and clean state per sessionPreview

Next-gen service with redesigned API format and runtime for building and operating agentsGA

Public Preview: Claude Opus 4.6, Available on Microsoft FoundryPreview