From Impressive Demos to Production AI Deployment
Microsoft Foundry is an enterprise AI platform that supplies runtime, tooling, and governance so that AI agents can move from experimental prototypes to dependable, production-grade systems that work with real data, real users, and real compliance requirements across an organization. Agentic AI has given engineering teams plenty of lively demos, but fewer systems that stay reliable under real load or satisfy security and audit expectations. At Build, Microsoft repositioned Foundry as an “AI app and agent factory,” presenting it as the place where enterprise AI agents move from experiments to production systems. Instead of adding only new model endpoints, Microsoft focused on the infrastructure layer enterprise developers previously had to assemble themselves: hosted runtimes, shared “memory,” knowledge retrieval, and policy-aware evaluation. Taken together, the updates signal that Microsoft now treats reliability, governability, and operational readiness as the main competitive advantage for enterprise AI agents, not raw model capability.
A Unified Runtime for Enterprise AI Agents
The centerpiece of the release is Foundry Agent Service, a managed runtime that hosts enterprise AI agents in sandboxed sessions with dedicated compute, memory, and filesystem access. Each session keeps its own state, and the same runtime supports long-running autonomous agents, short-lived chat interactions, and scheduled background work through routines now in public preview. Importantly for existing teams, agents built with Microsoft Agent Framework, GitHub Copilot SDK, LangGraph and other SDKs can be deployed without rewrites through a stateful Responses API or a lighter invocations protocol where teams control request and response formats. This directly addresses the gap between proof-of-concept and production AI deployment, where orchestration and state handling often become brittle homegrown code. By treating “memory” as a platform function — including procedural, user, and session memory — Foundry helps agents learn how to do work across runs, improving task success without custom memory plumbing.
Toolboxes, Skills, and Enterprise Data Without Custom Plumbing
As enterprise AI agents adopt more tools, services, and data sources, managing them one by one becomes a bottleneck. Toolboxes in Microsoft Foundry, now in public preview, offer a single managed endpoint for tools, skills, Model Context Protocol clients, and enterprise integrations, so agents discover what they need at runtime instead of having every tool wired in manually. Skills are treated as first-class, versioned assets in a project-scoped catalog and can be exposed as MCP resources, making reuse and lifecycle management less fragile. A tool search feature narrows down which tools an agent sees per task, which improves response quality and keeps context windows from bloating. Toolboxes connect directly to Microsoft IQ services, including Work IQ, Fabric IQ, and semantic models, so teams can plug enterprise data into AI agents without building custom connectors for each source. Foundry also supports publishing agents into Microsoft 365 Copilot and Teams, with identity and policy applied automatically.
AI Governance Platform: From Policies to Measurable Evaluations
Foundry’s governance additions are aimed at teams that must prove AI agents comply with internal rules as well as external regulations. Microsoft introduced ASSERT, an open-source framework for policy-driven evaluation and regression testing that converts written policies into concrete, measurable checks. Instead of relying only on static benchmarks, ASSERT generates targeted scenarios to reveal safety and quality defects before they reach production, and it works across frameworks such as LangChain, CrewAI, LightLLM, and OpenAI. According to InfoQ, the wider Foundry platform offers shared observability and policy across every agent, unifying how organizations monitor behavior, enforce constraints, and audit outcomes. Grounding and retrieval are addressed by Foundry IQ, which brings Work IQ, Fabric IQ, Azure SQL, file search, and web data behind a single retrieval endpoint with a shared service-level agreement. Together, these governance and knowledge layers push Foundry toward a full AI governance platform rather than a model hosting service.
Why Reliability and Operations Now Define Enterprise AI
For enterprise developers, the Foundry updates represent a shift from model-first experimentation to platform-level operations. Nick Brady wrote that the release brings “runtime, tools, memory, grounding, models, observability, and governance” that developers need for production agents, underscoring that Microsoft sees the competitive front in enterprise AI as reliability and governability, not capability. Hosted runtime, shared memory, and Foundry IQ reduce the amount of infrastructure teams must assemble. Toolboxes and skills catalogs organize tools and data in one place. ASSERT and shared policy enforcement help teams move from one-off red-teaming to continuous evaluation tied to written rules. Build’s positioning of Foundry as the unified platform for enterprise AI agent deployment suggests that successful AI programs will be defined less by the newest model and more by which platform makes it safe and predictable to run many agents at scale in production.
