From Demos to Production: What Microsoft Foundry Now Is
Microsoft Foundry is an enterprise AI platform that combines runtime, tooling, memory, grounding, models, observability, and governance so developers can move AI agents from experimental demos into reliable production-ready AI systems that run at real scale, on real data, under real compliance and policy constraints. At Build 2026, Microsoft described Foundry as an “AI app and agent factory” and “the place where AI agents move from experiments to production systems.” Instead of focusing on flashier model features, the release targets the missing infrastructure that teams have been assembling themselves: hosted runtimes, policy-aware evaluation, unified knowledge retrieval, and consistent AI agent governance. This shift signals that Microsoft sees the battle for enterprise AI infrastructure as being about reliability and control more than raw model capability, especially for organizations that must keep agents observable, auditable, and aligned with internal rules.
A Managed Runtime That Fits Existing Agent Frameworks
The new Foundry Agent Service gives Microsoft Foundry production agents a managed runtime that avoids rewrites. Every agent session runs in its own sandbox with dedicated compute, memory, and filesystem access, so long-running autonomous agents can hold durable state and files across runs. Hosted agents support two integration paths: a stateful Responses API compatible with OpenAI-style interactions, and a lighter invocations protocol for teams that control their own request and response formats. Routines, now in public preview, extend the same runtime for scheduled work such as overnight ticket triage or daily reporting, pushing agents toward background operational roles. Because agents built with Microsoft Agent Framework, GitHub Copilot SDK, LangGraph, and other SDKs can be deployed directly, enterprises do not have to abandon existing orchestration code to adopt the new runtime. Instead, Foundry quietly replaces the stitched-together infrastructure that previously surrounded many agent experiments.
Toolboxes and Memory: Infrastructure for Repeatable Enterprise Work
Toolboxes in Foundry, now in public preview, tackle the tool sprawl that appears as agents grow in capability. They give each agent a single managed endpoint for tools, skills, Model Context Protocol clients, and enterprise data integrations. Configure tools once, expose them as MCP resources, and Foundry enforces authentication, lifecycle management, and AI agent governance. Tool search helps select a focused set of tools per task, improving quality and keeping context windows from swelling. Foundry also treats memory as a platform concern rather than an application afterthought. Memory in Foundry Agent Service supports procedural, user, and session memory, with procedural memory teaching agents how to carry out work across runs. Nick Brady reports that early Tau benchmark results show “7 to 14 percent absolute success rate gains at near baseline cost” when procedural memory is enabled, pointing to measurable improvements in task completion for production-ready AI systems.
Governance, Grounding, and Microsoft’s Bid to Own the Agent Lifecycle
Governance and grounding features show how Microsoft aims to own the full lifecycle of enterprise AI infrastructure. ASSERT, an open-source evaluation framework, converts written policies into measurable tests and generates scenarios to surface safety and quality issues before deployment. It works across LangChain, CrewAI, LightLLM, OpenAI, and other frameworks, letting enterprises apply consistent policy-driven checks regardless of the agent stack. Foundry IQ provides a unified knowledge layer behind agents, spanning Work IQ, Fabric IQ, Azure SQL, file search, and other sources behind one SLA-backed retrieval endpoint. Serverless options and multi-source knowledge bases move retrieval closer to a standard service than a custom integration task. Together with direct publishing into Microsoft Teams and Microsoft 365 Copilot, these capabilities turn Foundry into a single control plane for building, grounding, evaluating, deploying, and observing agents—an attempt to make reliability, not raw model capability, the main competitive edge for Microsoft Foundry production agents.
