Browser storage is unavailable or a saved preference is invalid. The guide still works, but changes to saved items and progress may last only for this session.
Stay currentWhat changed?Search official announcements, GA milestones, previews, SDK releases, and retirements.
Find the right sourceWhere is the real how-to?Search official guides, reference implementations, samples, SDKs, and learning paths.
Make it practicalWhat should I build next?Follow a deployment, orchestration, gateway, knowledge, or evaluation runbook.
See it workCan I try a real deployment?Inspect the Sweden Central project and invoke deployed agents from the browser.
Your daily briefing
The changes worth your attention.
Actual upstream versions, distinct release channels, and dates you can trust. A quiet day is allowed; a stale refresh is not hidden.
Loading release channels...
Stable/prerelease describes the upstream package or GitHub release, not every feature's GA status. A prerelease is shown beside stable only when it was published more recently. CLI and runtime versions are distinct. Failed sources retain their last-known snapshots.
Make this your daily starting point
Loading your reading state...
Use Ctrl+K / Cmd+K to search everything. Save useful sources, share filtered URLs, and keep your place in a playbook. No account or tracking service required.
The resource library
Less hunting. More building.
A deliberately curated path through official documentation and runnable examples. Every card tells you what you will get, when the guide was reviewed, and whether its source has changed.
Loading the curated library...
From reading to running
Pick an outcome. Follow the runbook.
Practical learning sequences with official references, deployment decisions, and things to check before shipping. Progress is local to this browser; checking a box does not provision or validate Azure resources.
Loading playbooks...
Daily source refresh
What's new across Microsoft Foundry
Search product announcements, Python and .NET SDK releases, Foundry Local, evaluation packages, and watched documentation or sample changes. This is a bounded rolling archive, not an exhaustive product changelog.
!
A launch post is not a deployment guarantee
Check current Learn guidance for region, model, subscription, networking, and preview restrictions. SDK “stable”, documentation “updated”, and feature “GA” mean different things. See how the sources are tracked.
Loading latest updates…
Source of truth
Where this guide gets its facts
Not every source has the same authority. Product status follows Azure Updates and Microsoft Learn; blogs explain launches; repository feeds provide the most precise SDK and runtime details.
Loading source health…
The Big Picture
One platform, five pillars
Microsoft Foundry is an agent-first AI factory that spans cloud, edge, and Microsoft 365 — a shift from a model-centric toolbox to a complete platform for agentic apps. Everything below rolls up into these pillars.
01
Foundry Models
Unified catalog of frontier & open models — OpenAI, Anthropic, Grok, Mistral, Llama, Phi, Hugging Face — with routing, fine-tuning & provisioned throughput.
02
Foundry Agent Service
Managed runtime to build, host, orchestrate, and scale single- and multi-agent systems with identity, memory, and tools.
03
Foundry IQ & Tools
Knowledge and retrieval services plus governed tools that connect agents to data and actions. Check each connector's identity and data-access requirements.
04
Foundry Control Plane
Day 2 home base: observability, evaluations, red teaming, guardrails, policy, cost and fleet governance across every agent & model.
05
Foundry Local
On-device inference runtime for Windows, macOS & edge — run models and agents offline with the same tooling.
💡 Why it matters: Instead of stitching together a model host, an orchestration library, a vector DB, an eval harness, and a monitoring stack from five vendors, Foundry gives you the full lifecycle — build → deploy → operate → govern — under one roof, with Azure identity, networking, compliance, and billing already wired in.
Pillar 01 · Models
Foundry Models — choose the model and the deployment
The model catalog is the front door: sold-direct-by-Azure frontier models plus thousands of partner and open-weight models, all deployable behind a consistent API with unified auth, quota, and billing.
🗂️
Model catalog
Compare Microsoft, OpenAI, Anthropic, and other model providers in the current catalog. Model versions, modalities, terms, supported tools, and regional availability differ.
🔀
Model Router
Evaluate a routing deployment when workload quality and cost vary by request. Check its supported underlying models and API features; routing does not replace application-level evaluations.
🎛️
Fine-Tuning & RFT
Supervised fine-tuning, distillation, and reinforcement fine-tuning on frontier models; bring datasets, run training jobs, and deploy tuned variants under the same governance.
📦
Deployment choices
Compare Global Standard, Data Zone, regional, and provisioned options where the chosen model supports them. Throughput, data residency, quotas, billing, and SLAs must be checked per deployment.
🧩
Bring your own endpoint
An API Management AI gateway can put policies in front of supported model endpoints. Validate protocol compatibility, identity, monitoring, and routing rather than assuming all backends behave identically.
📊
Benchmarks & Leaderboards
Compare models on quality, cost, and throughput with built-in benchmarking before you commit — then A/B in production with evals.
Spotlight
Multi-model decisions without stale model tables
Use workload evidence, not a permanent “best model” label. Claude, OpenAI, Microsoft, and other model families have different APIs, hosting options, regional availability, prices, and limits. Open the current model documentation before committing.
| Workload | Compare | Measure | Do not assume |
|---|---|---|---|
| Complex reasoning | Higher-capability reasoning deployments | Task success, groundedness, latency, total reasoning tokens | A larger context window guarantees better answers |
| Tool-using assistant | Mid-tier models with the required tool/API support | Tool selection, structured outputs, handoff accuracy | Every provider supports the same API features |
| High-volume triage | Smaller models or a measured routing strategy | Classification quality, p95 latency, cost per resolved request | The cheapest token price means the cheapest workflow |
🏗️
Check the hosting contract
- Confirm the model provider, hosting location, supported deployment SKU, and residency terms.
- Check whether your app needs the OpenAI-compatible or provider-specific API and SDK.
- Confirm context/output limits, tools, streaming, structured outputs, and model retirement dates for the exact version.
💳
Make cost and governance explicit
- Read the current offer's billing, Marketplace, commitment, and data-processing terms.
- Measure input, output, reasoning, cache, and tool costs on representative workloads.
- Configure identity, logging/redaction, request budgets, and evaluation gates for the chosen model.
- Retest prompts and tools before changing the model version or routing policy.
Try the trade-off: the existing Azure demo includes concierge, analysis, and triage agents on different deployed model tiers. Use the live lab to compare behavior, and use an evaluation dataset before drawing quality or cost conclusions.
Pillar 02 · Agents
Foundry Agent Service — agents as managed cloud resources
The Agent Service turns "an agent" from a script on someone's laptop into a durable, secured, observable cloud resource: model + instructions + tools + knowledge + memory + identity, exposed as an API.
⚙️
Prompt agents
Version a definition containing instructions, model, and supported tools. Start here when you need a managed reasoning loop rather than your own hosted application code.
🧠
Memory and state
Choose conversation state, application storage, and supported persistent memory deliberately. Define consent, retention, tenant isolation, and what should never be remembered.
🛠️
Toolboxes and MCP
Connect supported tools through governed definitions. Validate each tool's authentication, permissions, approval requirements, and data boundary before giving the agent access.
🪪
Identity is configuration
Use Microsoft Entra and managed identities where supported. Assign minimum required roles explicitly; an agent or project existing does not prove its tools are authorized.
⏱️
Scheduled work
Use supported routines or an explicit scheduler for background tasks. Check current availability and put budgets, failure reporting, and cancellation around unattended work.
✨
Optimization
Improve agents against versioned datasets and evaluators. Check the current optimizer's access requirements; do not promote a candidate without holdout and safety results.
Spotlight
Hosted Agents — bring your code-first agents into Foundry
Bring your own agent code to a managed Foundry runtime. Start with the current direct-code deployment path when your application fits it; use a container when you need custom system dependencies or an existing image. Framework and protocol support still need to be checked.
1 · Build locally
Write your agent in Python/.NET with the framework of your choice.
→
2 · Choose packaging
Direct code + remote build, or an explicit container/ACR path.
→
3 · Deploy a version
Use the supported azd project and an existing or explicitly provisioned Foundry project.
→
4 · Operate
Invoke via the standard Responses API; scale, trace, evaluate & govern in the Control Plane.
💎
The value of bringing agents into Foundry
- Managed hosting: offload supported runtime management while owning code, dependencies, and behavior.
- Explicit security: configure identity, RBAC, and networking. Private access is not automatically inherited just by deploying an agent.
- Observable code: instrument the application and connect the appropriate telemetry and evaluation surfaces.
- Protocol choice: follow the agent's supported Responses, invocations, or real-time protocol rather than assuming one payload fits every agent.
- Versioned rollout: keep known-good code, configuration, evaluation results, and an explicit rollback path.
🧾
Start from a supported project, not a guessed manifest
The packaging contract and CLI evolve. Follow the current quickstart for azure.yaml, service configuration, runtime, entry point, dependencies, and invocation protocol.
- Prefer direct code for standard supported workloads.
- Use container deployment for genuine image/system requirements.
- Bind to the intended project and existing model deployment.
- Smoke-test, evaluate, and keep the previous version available.
Open Source SDK
Microsoft Agent Framework Python + .NET
An open-source framework for agents and explicit multi-agent workflows. Follow the separate Python and .NET release channels in the daily brief; read migration guidance before carrying forward Semantic Kernel, AutoGen, or older Agent Framework code.
🤖
Agents
LLM-driven, tool-using, autonomous. Ideal for open-ended tasks where the model decides the path.
- Model clients appropriate to the selected provider and API
- Function tools, MCP servers, OpenAPI, A2A messaging
- Application state and supported memory integrations
- Middleware for auth, filtering & telemetry
🧭
Workflows
Explicit, typed, graph-based orchestration for structured business processes where you decide the path.
- Patterns: sequential · concurrent · handoff · magentic (dynamic manager)
- Checkpointing & durable long-running state
- Human-in-the-loop pauses & approvals
- Type-safe routing with conditional edges
Start with the current client and migration guide
Install the package version you intend to support, follow the official quickstart for your language and provider, and run its minimal example before adding orchestration. Older AzureAIAgentClient snippets are not a safe substitute for the current API.
Production is a deliberate set of choices
- Configure OpenTelemetry and an exporter; review whether prompts and responses are captured
- Use Entra ID where supported; avoid keys in source code
- Open standards: MCP · A2A · OpenAPI for cross-vendor interop
- Adapt the application's hosting, state, identity, and protocol contract for Foundry hosted agents
- .NET: dotnet add package Microsoft.Agents.AI
Know the boundary: Agent Framework is your application framework; Foundry Agent Service is a managed service/runtime. You can combine them, but orchestration choices, hosting configuration, and service availability are separate decisions.
Pillar 03 · Knowledge
Foundry IQ — the shared knowledge layer
Foundry IQ (powered by Azure AI Search) turns retrieval into a reusable platform capability: knowledge bases that any agent can ground on, instead of bespoke RAG pipelines per app.
📚
Knowledge sources
Choose supported data sources and configure ingestion, indexing, synchronization, and retention. A connector existing does not establish your data's freshness or permissions.
🎯
Agentic retrieval
Use supported query planning and multi-source retrieval with citations. Compare retrieval quality and latency on your own dataset rather than assuming a universal relevance improvement.
🔒
Permission-aware retrieval
Configure supported document-level access controls and the caller identity flow. Test cross-user and cross-tenant boundaries explicitly; not every connector trims results the same way.
The production front door
AI gateway: where access, budgets, and routing meet.
Azure API Management can govern supported model, agent, and MCP traffic. Treat it as an explicit architectural layer, not a feature automatically enabled by creating a Foundry project.
01 · CallerYour application
Sign in with Entra. Request only the scopes and actions the user is entitled to use.
02 · PolicyAPI Management
Validate tokens, apply per-caller budgets, route requests, and govern tools.
03 · RuntimeFoundry + tools
Use managed identity or the documented identity flow for the selected backend.
04 · EvidenceTelemetry + evals
Correlate requests, latency, token usage, and quality without indiscriminately logging prompts.
Authenticate both boundaries
Caller-to-gateway and gateway-to-backend are different trust decisions. CORS is not authentication. Validate issuer, audience, permissions, and allowed callers; assign backend RBAC separately.
Budget before you retry
Choose supported token/rate policies per consumer and deployment. Handle 429s and backend pools deliberately. Retrying a stateful agent request can duplicate work or charges.
Check protocol and tool support
Test Responses, streaming, MCP, and agent semantics against the current gateway documentation. Do not apply a chat-completions policy blindly to every agent endpoint.
This site's live lab: requests currently go directly to the configured Foundry project with your short-lived Entra token. No APIM instance or anonymous paid-inference proxy is implied by this diagram. The gateway playbook is the path for adding that layer.
Pillar 04 · Operate & Govern
Day 2 Operations — where trust is earned
Shipping an agent is Day 1. Foundry's Control Plane makes Day 2 — evaluating, securing, watching, and governing AI in production — a continuous, automated discipline rather than an afterthought.
🧪
Evaluations
- Built-in evaluators: groundedness, relevance, coherence, fluency, safety, plus agentic evaluators — intent resolution, tool-call accuracy, task adherence
- Custom & LLM-as-judge evaluators for domain rules
- Run in dev, in CI/CD gates (GitHub Actions / Azure DevOps), and continuously on live traffic
- Batch evals over datasets curated straight from production traces
🛡️
AI Red Teaming Agent
- Automated adversarial testing built on Microsoft's open-source PyRIT framework
- Simulates prompt injection, jailbreaks, data exfiltration & harmful-content attacks across risk categories (violence, hate, self-harm, sexual)
- Reports Attack Success Rate (ASR) with per-attack logs for remediation
- Run pre-deployment or on a schedule against live agents for continuous regression checks
🔭
Observability & Tracing
- Instrument model calls, tool invocations, retrieval, and handoffs with supported OpenTelemetry integrations
- Unified dashboards in the Control Plane's Operate view — quality, cost, latency, token burn per agent
- Deep Azure Monitor / Application Insights integration and alerting on drift or quality regressions
- Set sampling, retention, sensitive-data capture, and access controls before using traces as incident evidence
🚧
Guardrails & Content Safety
- Azure AI Content Safety filters on inputs & outputs: hate, violence, sexual, self-harm
- Prompt Shields against injection & jailbreak attempts; groundedness detection against hallucination
- Protected-material detection and PII redaction
- Attach the supported safeguards and test their actual coverage across your application and tool boundaries
🏛️
Governance & Fleet Management
- Foundry Control Plane: one registry of every agent, model & tool with ownership and lifecycle state
- Entra Agent ID for identity & least-privilege permissions; conditional access for agents
- Integration with Microsoft Purview (data compliance) & Defender (threat protection)
- Maps to NIST AI RMF (Map · Measure · Manage) and supports EU AI Act readiness
💰
Cost & Quota Control
- Model/deployment quota plus explicit application or gateway limits for each consumer
- Spend dashboards spanning serverless, PTU, and Marketplace (e.g., Claude CCU) billing
- Model Router + prompt caching + batch API as built-in cost levers
A successful response is not an evaluation pass. Start with a versioned representative dataset, separate development from holdout cases, select supported quality and safety evaluators, and compare against a baseline. Set thresholds for your workload; log evaluator and model versions alongside the agent version.
| Evidence to collect | What it answers | Use it for |
|---|---|---|
| Functional smoke | Did authentication and invocation return a completed, nonempty response? | Deployment readiness, not proof of answer quality |
| Offline quality + safety | Does this version meet the rubric across representative and adversarial cases? | Release gates and baseline comparison |
| Production traces + sampling | Where do tool failures, latency, cost, and quality regressions occur? | Operations, targeted datasets, and continuous evaluation |
| Holdout + rollback record | Did the improvement generalize, and can you restore the known-good version? | Safe promotion rather than optimizing to the test set |
Evaluation & observability runbookEvaluator, tracing, and safety referencesUnderstand this demo's validation boundary
The GenAIOps loop
Build
Pick models, define agents, ground with IQ
→
Evaluate
Quality + safety evals, red team scans
→
Deploy
CI/CD gates → hosted agents / endpoints
→
Observe
Traces, dashboards, alerts, cost
→
Improve
Trace-driven datasets, prompt optimization, re-eval
Pillar 05 · Edge & Ecosystem
Foundry Local & the broader ecosystem
💻
Foundry Local
Run supported models on-device using the current runtime and SDK. The v2 release line introduces a Session API; older OpenAI-style in-process examples need migration. Check OS/hardware support, download requirements, and model licensing before promising offline operation.
🧑💼
Microsoft 365 & Teams
Explore supported agent publishing flows and check current access, tenant, and channel restrictions. Availability of the agent runtime does not establish availability of every publishing destination.
👩💻
Developer tools
Use the Foundry portal, supported VS Code tooling, Azure Developer CLI (azd), and the SDK/API appropriate to your agent. Keep tooling, project configuration, and sample versions aligned.
Cheat sheet — what to reach for, when
| You want to… | Use |
|---|---|
| Call a frontier model (GPT-5, Claude, Grok…) with enterprise auth | Foundry Models — serverless deployment or PTU |
| Auto-pick the cheapest model that's good enough | Model Router |
| Stand up a tool-using agent without writing infra | Foundry Agent Service (prompt agents) |
| Run your LangGraph / CrewAI / MAF agent as a managed service | Hosted agents (container → Foundry) |
| Write agent code with orchestration, memory & HITL | Microsoft Agent Framework (Python / .NET) |
| Ground agents on enterprise data with permission trimming | Foundry IQ knowledge bases |
| Prove quality before & after shipping | Evaluations (+ CI/CD gates) |
| Attack your own app before adversaries do | AI Red Teaming Agent (PyRIT) |
| Watch, alert, and audit production agents | Control Plane observability + Azure Monitor |
| Block harmful content & prompt injection | Content Safety guardrails + Prompt Shields |
| Give every agent an identity & least privilege | Entra Agent ID |
| Run models offline / on-device | Foundry Local |
Your Azure learning lab
Read the guide. Put it to work.
Use a real versioned agent in your existing Sweden Central project to walk through the runbooks. Load a scenario, attach the guide's dated official-source context, and inspect the API request and response. Nothing runs in Azure just because you open this page.
Foundry showcase
Not connected
Connect to load the current agent inventory.
Get a short-lived token with Azure CLI
az account get-access-token --scope https://ai.azure.com/.default --query accessToken -o tsvScenario workbench
Loading scenarios...
Inspect the token-free request
Connect to view a token-free example request.