From Agent Demos to Enterprise AI Infrastructure
Microsoft Foundry is an enterprise AI infrastructure platform that provides a managed runtime, shared tools, and governance controls so organizations can move AI agents from experimental demos into scalable, compliant production systems across their existing development frameworks and business applications. At Build 2026, Microsoft shifted the story of Foundry from a developer playground to what it calls an “AI app and agent factory,” aimed at teams that have been wiring their own orchestration, policy, and observability layers. Instead of selling another model endpoint, Microsoft is selling the glue: hosted agents, shared memory, grounded retrieval, consistent evaluation, and policy-driven governance. Nick Brady, senior program manager for developer experience at Microsoft, wrote that the release adds “runtime, tools, memory, grounding, models, observability, and governance” that production AI agents need, signaling that reliability and oversight, not raw model power, are becoming the main competitive front for enterprise AI platforms.
Hosted Agent Runtime Without a Rewrite
The centerpiece is Foundry Agent Service, a hosted runtime that turns prototypes into production AI agents without forcing teams to rewrite them. Each agent session gets its own sandbox with dedicated compute, memory, and a durable filesystem, so long-running flows and multi-step processes survive failures and restarts. Critically, the service supports agents built with Microsoft Agent Framework, GitHub Copilot SDK, LangGraph, and other SDKs, so existing work can move straight into production. Two protocols shape how teams integrate: a stateful Responses API compatible with OpenAI-style interactions, and a lighter invocations protocol for pass-through calls where teams control the payload. The same runtime powers scheduled routines in public preview, enabling agents to run on timers for jobs such as overnight ticket triage or daily reporting. This design makes Foundry feel less like another model hub and more like an application runtime for production AI agents.
Toolboxes and the New AI Agent Governance Layer
As AI agents gain more tools, the challenge becomes AI agent governance: deciding which tools each agent may call, under what policies, and with what audit trail. Foundry’s Toolboxes, now in public preview, turn that sprawl into a single managed endpoint per agent. Teams register tools, skills, Model Context Protocol clients, and enterprise integrations once; agents then discover them at runtime instead of wiring them in manually. Skills are versioned in a project-scoped catalog and exposed as MCP resources, while a tool search feature selects a small set of relevant tools per task instead of dumping entire catalogs into the context window. Toolboxes also connect agents to Microsoft IQ services such as Work IQ, Fabric IQ with the Fabric data agent, Ontology, and semantic models, so teams can attach enterprise data sources without custom pipelines for every agent or framework.
Memory, Grounding, and Observability as Platform Features
Foundry treats durable memory and grounding as shared platform services instead of one-off application code. Memory in Foundry Agent Service now spans user, session, and procedural memory, with the last one introduced at Build to help agents learn how to do work across runs. According to Nick Brady, early Tau bench results show “7 to 14 percent absolute success rate gains at near baseline cost” when procedural memory is enabled. The platform automatically extracts key facts and procedures from interactions, consolidates them by identifiers like Entra ID, and makes them inspectable and controllable for compliance teams. For knowledge, Foundry IQ unifies Work IQ, Fabric IQ, Azure SQL, file search, and other sources behind a single retrieval endpoint with a shared SLA. New options such as Foundry IQ Serverless and multi source knowledge bases aim to standardize how production AI agents access and ground against enterprise data.
ASSERT, Agent Control Spec, and Pushing Agents to Where Work Happens
Governance is where Microsoft is betting Foundry will stand out among enterprise AI infrastructure platforms. ASSERT, an open-source framework for Adaptive Spec-driven Scoring for Evaluation and Regression Testing, converts written policies into measurable evaluations and generates targeted scenarios so teams can detect safety and quality issues before deployment. It already supports frameworks including LangChain, CrewAI, LightLLM, and OpenAI. Microsoft is also introducing Agent Control Spec, an open specification for describing what agents are allowed to do and how they should be controlled, so organizations can standardize policies across teams and tools. On the distribution side, Foundry adds direct publishing into Microsoft Teams and Microsoft 365 Copilot, planned for general availability in June 2026. That step lets production AI agents built on the Microsoft Foundry platform appear directly in employee workflows, with identity, permissions, and policy applied automatically.
