Frontier models and production agents: Advancing Microsoft Foundry for the agentic era

AI reaches its potential when it’s integrated into reliable and observable systems that actively contribute to business success. Currently, over 100,000 organisations are leveraging Microsoft Foundry, with notable companies like Adobe, Telefónica, and Tata Consultancy Services actively using agents in production.

During Microsoft Build, we emphasised a straightforward promise for the agentic era: developers should be able to create agents in their current environments, run them on trusted infrastructure, and present them to users efficiently—without assembling disjointed platforms. This ambition is becoming a reality today, with three crucial updates now generally available in Microsoft Foundry:

• OpenAI’s latest frontier model series: Featuring GPT-5.6 Sol, GPT-5.6 Terra, and GPT-5.6—each tailored for different workloads and available in Standard Global and Standard Data Zones.
• Asia-Pacific Data Zone: Offering customers in the APAC region a local option to run frontier OpenAI models, ensuring data processing remains within their geographical region.
• Production agents in Foundry Agent Service, which includes hosted agents, toolboxes, and the ability to publish to Microsoft 365 Copilot and Microsoft Teams.

These updates converge advanced models, agent runtimes, enterprise-level security, and compliance controls into a unified platform—enabling organisations to transition from testing to production seamlessly without needing multiple disconnected tools.

Why Foundry is the premier platform for agents

Microsoft Foundry is designed as a comprehensive platform for building, running, managing, and distributing AI agents. It provides the essential tools organisations need to take agents from concept to production through three main pillars:

• Build: Open and adaptable across various models and frameworks.
• Generate: Seamlessly connected to enterprise data, tools, and users.
• Govern: Secured, supervised, and optimised for lasting value.

These elements form the foundation of the recent Foundry updates, facilitating the building, running, and scaling of production agents all within one cohesive platform.

Develop with any framework on the industry’s AI platform

Development begins seamlessly where developers are most comfortable—be it in GitHub Copilot or Microsoft Visual Studio (VS) Code. The Foundry Toolkit for VS Code and the Foundry skill handle deployments to Foundry. Whether you’re using the Microsoft Agent Framework, the GitHub Copilot SDK (now readily accessible), or the Claude Agent SDK, Foundry serves as the ultimate production environment—beginning with the most suitable model.

Choose the right model for the task

An agent’s capabilities depend on the underlying model. Microsoft Foundry provides access to top-tier frontier, open-source, and task-specific models all in one place, allowing teams to select the ideal model for each workload.

We’re excited to announce that OpenAI’s GPT-5.6 series is now generally accessible in Microsoft Foundry Models and Microsoft Foundry Agent Service:

• GPT-5.6 Sol: Offers the most advanced reasoning skills, perfect for detailed analysis, automated workflows, and coding tasks for heavy enterprise demands.
• GPT-5.6 Terra: A well-balanced model suitable for regular tasks, delivering performance comparable to GPT-5.5 at a lower cost—ideal for widespread intelligent application deployment.
• GPT-5.6 Luna: The fastest and most affordable model, perfect for high-traffic and time-sensitive tasks.

Together, the GPT-5.6 series allows organisations to align model capabilities, costs, and performance directly with specific business needs, avoiding the pitfalls of using a single model for all workloads.

Clients consistently express the importance of having access to new models alongside quality. Hence, we are rolling out GPT-5.6 through Global Standard and Global Priority Processing in all existing 28 global regions, Data Zones Standard, and Global Provisioned from day one. This means customers can integrate the latest AI developments exactly where they’ve already built, deployed, and expanded their applications.

Pricing for GPT-5.6

Implement frontier AI in your operating region

Introducing more models in additional regions is just part of the expansion. The other significant aspect is the ability to run these models compliantly. This is where the Asia-Pacific Data Zone comes into play. We are pleased to announce the general availability of the Asia-Pacific (APAC) Data Zone for Microsoft Foundry, enabling APAC customers to deploy frontier OpenAI models while ensuring data processing remains within their region, without having to create a separate environment.

With Global, Data Zone, and Regional deployment options, organisations can ensure their AI practices meet sovereignty, compliance, performance, and scalability requirements, maintaining a consistent development and operational experience across various environments.

As financial institutions embrace AI, responsible data handling becomes essential for building trust. Microsoft Foundry’s APAC Data Zone permits us to keep data processing locally while accessing cutting-edge AI models at scale. This gives us the confidence to drive AI innovation responsibly, reinforcing our goal to be a leading AI-powered financial platform in Asia.

—Hongsoo Kim, Chief Data and AI Officer (CDAO), Viva Republica (Toss)

Create meaningful impact with context-aware, action-driven agents

While a capable model is essential, it’s just the beginning. For an agent to be effective in production, it must run in a robust environment, understand your business context, have secure access to necessary tools, remember past interactions, and effectively engage with users. Foundry offers all of these capabilities as built-in features designed to work seamlessly together.

• Where it operates: Hosted agents in Foundry Agent Service are now generally available, providing developers with a unified runtime for agents developed with any framework—be it Microsoft Agent Framework, GitHub Copilot SDK, LangGraph, OpenClaw, or Hermes. It’s built for enterprise use from day one: Network isolation with Microsoft Azure Virtual Network (VNet) integration ensures agent traffic remains secure. For extended workloads, the new resilient task support in hosted agents (available in private preview) makes it easier to create agents capable of recovering from failures. The platform will manage continuous operation, allowing for multi-turn conversations and human-in-the-loop approvals to resume smoothly after interruptions.
• How it communicates: Hosted agents with Voice Live is now available for developers to incorporate real-time voice capabilities into their agents using the Azure VoiceLive SDK.
• What it understands: Foundry provides agents with access to enterprise knowledge, eliminating the need for developers to build complex retrieval systems. Microsoft IQ integrates Work IQ for real-time insights into your Microsoft 365 ecosystem, Fabric IQ for structured data handling, and Web IQ for real-time web integration—all unified under Foundry IQ, now generally available as the dependable knowledge framework behind every Foundry agent.
• How it accesses tools: Toolboxes in Foundry are now available. These allow agents to dynamically select the appropriate tool, providing curated, secure access while significantly reducing unnecessary token usage in large tool sets.
• How it remembers and reacts: Memory androutines in Foundry Agent Service are currently in public preview. Memory (procedural, user, and session-based) allows agents to retain context across different interactions. Routines enable agents to operate on a schedule or respond to real-world events, such as a ticket being filed or a workflow being completed, automatically reacting to changes.
• How it engages with users: Publishing to Microsoft Teams and Microsoft 365 Copilot will be generally available next week. This allows the agents your developers create to integrate directly into applications used by countless professionals, with identity and permissions being managed automatically. This is effective even for network-isolated agents: if a project operates behind a private endpoint, publishing can happen through a documented flow instead of a one-click solution, keeping the agent secured while Microsoft’s channel adapters connect via your own firewall.

Manage and enhance the entire AI lifecycle with oversight and controls

An agent that isn’t visible, improvable, or secure isn’t suitable for real-world application—therefore, Foundry prioritises trust as a platform essential rather than placing the burden solely on developers. This latest release enhances monitoring and post-deployment improvements so you can evaluate agent performance and value.

• How you track performance: Tracing andevaluation tools for hosted agents are fully available. This lets you see precisely what an agent accomplished, the rationale behind its actions, and where any issues arose, allowing for systematic evaluations pre- and post-deployment.
• How to enhance efficiency: The agent optimizer in Foundry Agent Service, currently in public preview, can test your prompts, skills, models, and tools collectively, automatically identifying optimisation avenues—often allowing you to maintain quality standards while transitioning to smaller, more cost-effective models.
• How to demonstrate value: ROI for agents in Microsoft Foundry, currently in private preview, connects an agent’s tracking data, business evaluation metrics, and operational expenses into a streamlined view, highlighting KPIs like net value and total costs. This transparency allows teams to assess whether a production agent offers more value than the costs associated with its operation, and provides insight into tracing when value diminishes.

As agents transition from pilot testing to performing thousands of actions daily, Foundry equips teams with necessary tools to keep budgeting on track without switching platforms.

It all begins with choice.

• Select your deployment method: Foundry offers Global, Data Zone, and Regional options, allowing you to align AI practices with sovereignty, compliance, and performance demands while running frontier models with in-region data processing.
• Choose your model pricing: A complete range of options—Standard, Priority Processing, Provisioned Throughput, and Batch—enables you to optimise for agility, speed, volume, and costs on a single platform.

On this strong foundation, the model router ensures each request is matched to the appropriate model, prompt caching reduces unnecessary computations, and PTU spillover and quota optimisation ensure service remains stable during usage spikes. For agents, toolboxes in Foundry send only the required tools for each request, while the agent optimizer refines prompts, skills, tools, and model choices in accordance with your own evaluators.

Managing costs is just part of the picture. The ROI for agents in Foundry connects business value and usage costs in one view, so teams can identify when a production agent generates more value than it incurs expenses and pinpoint areas where costs exceed value.

For a practical overview, check out our latest Microsoft Mechanics episode focused on token economics for agents.

Real-world applications: What teams are creating using Foundry

The organisations utilising Foundry are not merely experimenting; they are actively shipping—ranging from digital innovators to leading global enterprises.

• Adobe is leveraging GitHub, Foundry Agent Service, and Azure Functions to deploy agents for their applications, significantly streamlining their time to production.
• Telefónica has embraced Microsoft Foundry as the backbone of their corporate agent platform, with agents tackling complex network operations—a major focus for any telecommunications company—across Microsoft Agent Framework, hosted agents, AI Gateway, and Azure Logic Apps.
• Tata Consultancy Services is implementing the agent optimizer within Foundry Agent Service to enhance agent efficiency through a structured approach to tuning prompts, which simplifies manual tasks while improving adherence and execution speed.

The trend is clear: teams that once took weeks to assemble, secure, and deploy agents can now transition these processes to days, using infrastructure that satisfies their compliance standards, while engaging users through familiar tools.

How to get started

Everything discussed here is currently live in Microsoft Foundry.

Check out the official documentation and Microsoft Learn courses to get started. Developers can jump in quickly by following the Quickstart, providing a step-by-step guide to setting up, testing, and deploying a production-ready hosted agent.

Explore AI Agents for Beginners for a comprehensive 12-lesson curriculum, or go deeper with structured labs like Develop AI Agents in Azure, Hosted Agents Workshop (.NET), the Foundry Toolkit for VS Code and hosted agents workshop, and the ZavaShop Supply Chain Workshop. For effective agent quality management, refer to Evaluating AI Agents: A Practical Guide with Microsoft Foundry.

Watch: Foundry Agent Service + Microsoft Agent Framework Explained—Jeff Hollan walks you through how to operationalise AI agents from deployment to real-world impact.