Microsoft Foundry Enhances Agent Production with Runtime, Tooling, and Governance

Microsoft has unveiled a suite of new functionalities for Microsoft Foundry, its AI application and agent development platform. Announced at the Build 2026 event, these enhancements are designed to equip developers with the necessary tools to transition AI agents from experimental phases into fully operational production systems. The updates focus on critical areas including runtime environments, development tooling, memory management, grounding capabilities, model integration, observability, and governance.

Foundry, described by Microsoft as an “AI app and agent factory,” is presented as a unified Azure platform. Its interoperable nature aims to enable teams to build, ground, and govern AI applications and agents that possess a deep understanding of business context, while maintaining shared observability and policy enforcement across all agents.

Enhanced Agent Runtime

The new release brings significant improvements to the Foundry Agent Service, which now offers managed, sandboxed sessions for hosted agents. These sessions provide state persistence, file system access, and support for various AI frameworks. Developers can leverage a stateful Responses API for detailed interactions or a lighter weight invocations protocol for direct passthrough calls. This robust runtime environment is capable of supporting long-running agents, such as OpenClaw and Hermes, ensuring durable state and file management. Additionally, routines, currently in public preview, allow for scheduled agent execution, facilitating tasks like overnight data processing or daily report generation. These additions build upon the general availability release of the Foundry Agent Service in 2025, which introduced capabilities like multi-agent orchestration, Agent-to-Agent APIs, and support for popular frameworks like Semantic Kernel, AutoGen, and CrewAI.

Streamlined Tooling and Distribution

Foundry’s new Toolboxes, now in public preview, offer a centralized, managed endpoint for agents to access tools, skills, Model Context Protocol clients, and enterprise data integrations. This approach allows tools to be registered once and discovered at runtime, rather than being hardcoded into each individual agent. Skills can be versioned and scoped to specific projects, exposed via MCP, and tool search functionality helps the platform identify the most relevant tools for a given task, avoiding the need to expose all available tools to the model.

A key aspect of the distribution enhancements includes direct publishing from Foundry into Microsoft Teams and Microsoft 365 Copilot. Planned for general availability in June 2026, this feature will allow agents developed in Foundry to be seamlessly integrated into the workflows where employees already operate, with automatic application of identity, permissions, and policies.

Advanced Memory and Grounding Capabilities

Microsoft Foundry treats “memory” as a core platform function rather than an application-specific feature. The Foundry Agent Service’s memory capabilities, introduced in public preview in late 2025, now encompass procedural, user, and session memory. Procedural memory, a new addition from Build, is designed to assist agents in learning and executing tasks across multiple runs. Early benchmarks indicate improved task success rates when this feature is enabled. Previous coverage of Foundry memory detailed how the service extracts and consolidates key information and procedures from conversations, storing them in a managed repository scoped by identifiers such as Entra ID, with controls for data retention and inspection.

Grounding and retrieval are addressed through Foundry IQ, which acts as a knowledge layer behind agents. It unifies data from various sources, including Work IQ, Fabric IQ, Azure SQL, and file search, providing a single, SLA-backed retrieval endpoint. At Build, Microsoft announced Foundry IQ Serverless in public preview and multi-source knowledge bases in general availability. Microsoft Web IQ was also introduced for live web grounding, promising sub-200 millisecond responses and zero data retention guarantees, alongside enhanced security features like encryption, permission synchronization, and sensitivity label governance.

Model Integration and Fine-Tuning

Foundry’s model catalogue has been expanded with four first-party MAI models in public preview: MAI Thinking 1 for chat and reasoning, MAI Image 2.5 for image generation and editing, MAI Transcribe 2 for speech-to-text with diarization, and MAI Voice 2 for multilingual text-to-speech with voice cloning. Furthermore, Fireworks AI on Foundry is now generally available, providing access to open models through a single Azure endpoint with enterprise SLAs, support for custom weight models, and integration with Foundry’s access controls and logging.

Managed Compute in Foundry Models is designed to optimize workload routing around regional GPU constraints. It supports fine-tuning and a capability called ‘Frontier Tuning,’ which Microsoft claims offers significantly greater cost efficiency compared to using models like GPT 5.5 directly for tasks such as generating technical documentation.

Foundry as a Production Platform

The layered approach to AI development is further clarified by comparing Microsoft 365 Copilot Agent Builder and Copilot Studio with Foundry. While Agent Builder and Copilot Studio offer low-code, visual experiences, Foundry is positioned as the code-first platform for advanced development, evaluation, and observability. Foundry is recommended for teams requiring custom logic, sophisticated retrieval mechanisms, and deeper integration with developer workflows. Microsoft’s guidance for secure agent development emphasizes mapping agent interactions with existing build, test, and release processes, applying the same discipline used for microservices, including clear scope definition, policy enforcement, tracing, and continuous evaluation—features now supported as first-class capabilities within Foundry. This evolution leads to Foundry being perceived as a robust production platform for agents, moving beyond its initial role of wiring up demos.

Datos clave
| Feature | Status | Key Benefit |
|—|—|—|
| Foundry Agent Service Runtime | Enhanced | Managed sessions, state persistence, file access for production agents. |
| Toolboxes | Public Preview | Centralized tool access, versioning, and discovery for agents. |
| Foundry IQ | General Availability | Unified knowledge layer for grounding across multiple data sources. |
| MAI Models | Public Preview | New first-party models for chat, image, speech, and voice functionalities. |

These advancements in Microsoft Foundry are crucial for organizations looking to operationalize AI agents effectively. By providing a cohesive set of tools for development, deployment, and management, Microsoft is enabling teams to build more reliable, scalable, and governable AI solutions that can be integrated directly into existing business processes and workflows.

Fuente: infoq.com – https://www.infoq.com/news/2026/06/microsoft-foundry-agents/

Datos clave