From Agent Demos to Production AI Agents
Microsoft Foundry is Microsoft’s unified enterprise AI platform for building, grounding, and governing AI agents that move from experimental demos to reliable production systems with shared observability and policy. At Build, Microsoft framed Foundry as “the place where AI agents move from experiments to production systems,” positioning it as an AI app and agent factory rather than a collection of model endpoints. This shift answers a growing problem: impressive agentic AI demos rarely survive real workloads, compliance checks, or messy enterprise data. Microsoft Foundry agents now sit on top of a managed infrastructure layer that connects workflows, tools, memory, and knowledge services into a single experience. The focus is less on adding the most capable model and more on making agents dependable, observable, and governable across every business unit that adopts them.
A Managed Runtime for Enterprise AI Infrastructure
The new Foundry Agent Service functions as an enterprise-grade runtime for production AI agents, with each agent session running in its own sandboxed environment with dedicated compute, memory, and durable filesystem access. Microsoft emphasizes that teams do not have to rewrite agents to benefit: agents built with Microsoft Agent Framework, GitHub Copilot SDK, LangGraph, and other SDKs can be deployed as-is. Two access patterns support different integration styles. A stateful Responses API offers an OpenAI-compatible interface, while an invocations protocol supports passthrough scenarios where teams manage their own request and response formats. Routines, now in public preview, let agents run on schedules for tasks like overnight ticket triage or daily reporting, backed by durable state for long-running autonomous agents. This runtime turns Foundry into a core piece of enterprise AI infrastructure, not a prototype playground.
Toolboxes and AI Agent Governance for Tools and Data
As enterprises add more tools to their AI stack, tool management becomes a governance problem. Toolboxes in Microsoft Foundry, now in public preview, give each agent a single managed endpoint that aggregates tools, skills, Model Context Protocol clients, and enterprise data integrations. Configure tools and skills once, point any MCP-compatible client at a single URL, and Foundry handles authentication, lifecycle, and AI agent governance policies centrally. Skills are versioned in a project-scoped catalog and exposed as MCP resources, while tool search helps Foundry select a small, relevant set of tools for each task instead of pushing an entire catalog into the model’s context window. Toolboxes also connect to Microsoft IQ services such as Work IQ, Fabric IQ with the Fabric data agent, Ontology, and semantic models, so Microsoft Foundry agents can reach enterprise data without custom plumbing for every source.
Memory, Grounding, and Knowledge as Platform Features
Microsoft treats memory and grounding as platform services rather than features every team must rebuild. Memory in Foundry Agent Service, available in public preview, offers procedural, user, and session memory. Procedural memory is designed to help agents learn how to carry out 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 it is enabled. Foundry IQ provides a unified knowledge layer that sits behind Microsoft Foundry agents, abstracting sources such as Work IQ, Fabric IQ, Azure SQL, file search, and others under a single SLA-backed retrieval endpoint. At Build, Microsoft announced Foundry IQ Serverless in public preview and multi source knowledge bases in general availability, along with Microsoft Web IQ for live web content. Together, these services give enterprises consistent grounding and retrieval behavior across every production AI agent.
Governance, Evaluation, and Distribution as Competitive Edge
Governance and operational maturity are where Microsoft believes the next enterprise AI competition will be decided. ASSERT, Microsoft’s open-source framework for policy-driven evaluation, translates written policies into concrete tests and generates targeted scenarios to find safety and quality defects before AI agents reach production. It works across frameworks such as LangChain, CrewAI, LightLLM, and OpenAI, making it easier to apply consistent governance to heterogeneous stacks. Foundry’s shared observability and policy model extend across every agent, and direct publishing from Foundry into Microsoft Teams and Microsoft 365 Copilot is planned for general availability, so agents can appear where employees already work with identity, permissions, and policy applied automatically. In this model, Microsoft Foundry positions reliability, observability, and AI agent governance as the real differentiators, shifting attention away from raw model horsepower toward production-ready, enterprise-scale AI systems.
