From Demos to Durable AI Agents
Microsoft Foundry is Microsoft’s enterprise AI platform that turns experimental AI agents into production systems by providing a managed runtime, shared tooling, and policy-driven governance that sit between models and business workloads so teams can deploy at scale with observability, control, and repeatable reliability. The wave of agentic AI has produced eye-catching demos, but far fewer agents that survive real traffic, live data, and compliance rules. At its Build conference, Microsoft framed Foundry as the missing infrastructure layer for AI agent production deployment, filling the gap between proof-of-concept bots and hardened applications. Instead of focusing on the newest model tricks, the platform now centers on runtime guarantees, state handling, and policy enforcement. This signals a strategy shift: Microsoft is betting that in enterprise AI infrastructure, the deciding factor will be reliability, governability, and operational maturity rather than marginal gains in model capability.
Open Runtime, Proprietary Control Plane
The clearest signal of that strategy is the decision to make the agent runtime free and open-source, through projects such as Scout running on OpenClaw, while keeping the control plane and governance stack as the main product. Enterprises can run the same execution engine on their own infrastructure, but Foundry’s value lies in how runs are observed, governed, and connected to corporate systems. Hosted agents in Foundry Agent Service provide managed sandboxes with dedicated compute, memory, and durable filesystem access, and can host long-running agents like OpenClaw and Hermes with persistent state. According to Microsoft’s Nick Brady, Foundry now adds “runtime, tools, memory, grounding, models, observability, and governance” rather than only new models. This split makes the runtime a common foundation while positioning the Microsoft Foundry platform as the place where policies, auditing, and fleet-wide management are applied.
Runtime and Tooling Built for Production Workloads
At the runtime layer, Foundry focuses on interoperability and operations. Agents built with Microsoft Agent Framework, GitHub Copilot SDK, LangGraph, and other SDKs can be deployed without rewrites. Two protocols are supported: a stateful Responses API that feels like OpenAI-style chat, and an invocations protocol for passthrough calls where teams control request and response formats. The same runtime powers routines in public preview, so agents can run scheduled jobs such as overnight ticket triage or daily reporting with durable state. On the developer side, Foundry Toolkit for VS Code is now generally available, letting teams create agents from templates, use GitHub Copilot for coding, debug with trace visualization, and deploy directly to Foundry Agent Service. Direct publishing into Microsoft Teams and Microsoft 365 Copilot means AI agent production deployment can reach end users with identity, permissions, and policy applied consistently.
Toolboxes, Memory, and the Knowledge Layer
Foundry’s Toolboxes address the growing problem of tool governance in complex agent systems. Instead of wiring tools into each agent, Toolboxes expose a single managed endpoint for tools, skills, Model Context Protocol clients, and enterprise data connections. Skills are cataloged and versioned, discoverable as MCP resources, while tool search helps each task receive only a focused set of tools, keeping context lean and quality higher. Foundry treats memory as a platform concern: its Agent Service offers user, session, and new procedural memory that helps agents learn how to perform work across runs. Procedural memory has shown 7 to 14 percent absolute task success gains at near baseline cost when enabled. Behind all of this sits Foundry IQ, a knowledge layer that unifies Work IQ, Fabric IQ, Azure SQL, files, and web data behind a single retrieval endpoint with a shared service-level agreement.
Governance as the Differentiator in Enterprise AI
Governance features make Microsoft’s bet explicit: enterprise AI battles will be won on reliability and control, not raw performance. ASSERT, Microsoft’s open-source Adaptive Spec-driven Scoring for Evaluation and Regression Testing, turns written policies into concrete evaluation criteria and generates targeted scenarios so teams can catch safety and quality defects before agents reach production. It works across frameworks including LangChain, CrewAI, LightLLM, and OpenAI-based agents, aligning testing with enterprise policy rather than static benchmarks. The Agent Control Spec adds a shared, open specification for defining how agents are configured, audited, and constrained in different environments. Together with hosted runtime, Toolboxes, and the knowledge and memory layers, these capabilities position the Microsoft Foundry platform as a unified enterprise AI infrastructure stack, aimed squarely at organizations that have working prototypes but lack a reliable way to deploy, monitor, and govern AI agents at scale.
