From Impressive Demos to Reliable AI Agents in Production
Microsoft Foundry is an enterprise platform that turns experimental AI agents into production-ready systems by providing a managed runtime, shared memory, knowledge retrieval, observability, and governance so that teams can deploy, monitor, and control agentic workflows across their existing application and data landscape. While the agentic AI wave has produced colorful demos, many agents still struggle under real load, complex data, and compliance constraints. At its Build event, Microsoft repositioned Foundry as “the place where AI agents move from experiments to production systems,” signaling that the main battle in enterprise AI agents production is now about reliability, not raw model capability. Instead of only adding new model endpoints, this release focuses on the surrounding platform: runtime isolation, policy-based evaluation, tool governance, and native integration with Azure services and Microsoft 365, all delivered as a managed experience for development teams.
Hosted Agent Runtime: Sandboxed, Durable, and Framework-Agnostic
The new Foundry Agent Service runtime is the backbone of Microsoft Foundry’s production story. Each agent session runs in its own sandbox with dedicated compute, memory, and durable filesystem access, so long-running or high-volume workloads do not interfere with one another. Agents built on Microsoft Agent Framework, GitHub Copilot SDK, LangGraph, and other SDKs can be deployed without rewrites, via a stateful Responses API or a more flexible invocations protocol where teams design their own request and response formats. This matters for enterprises that already maintain custom orchestration or multi-agent pipelines. Routines, now in public preview, allow agents to run on schedules for tasks such as overnight ticket triage or daily reporting, while durable state supports autonomous agents that span many interactions. Together, these runtime features turn Foundry from a developer preview into something closer to enterprise-grade infrastructure for agentic workflow automation.
Toolboxes, Memory, and Knowledge: Building Reliable Agentic Workflows
Tooling in Microsoft Foundry aims to tame the growing complexity of agent tools and data integrations. Toolboxes, in public preview, give agents a single managed endpoint where tools, skills, Model Context Protocol clients, and enterprise data connectors are registered once and discovered at runtime. Skills are versioned in a project-scoped catalog and exposed as MCP resources, while tool search narrows each task to a small, relevant tool set to protect quality and context windows. Foundry treats memory as a platform capability, not an app add-on: procedural, user, and session memory help agents learn how to perform work across runs. According to Nick Brady, early Tau benchmark results show procedural memory delivering “7 to 14 percent absolute success rate gains at near baseline cost.” Foundry IQ then unifies Work IQ, Fabric IQ, Azure SQL, and file search behind one retrieval endpoint, so agents tap enterprise data without custom plumbing.
Governance, Evaluation, and Enterprise-Ready Distribution
Enterprise AI governance is central to this Foundry release. ASSERT, Microsoft’s open-source Adaptive Spec-driven Scoring framework, evaluates agents against written policies instead of static benchmarks, generating targeted scenarios to surface safety and quality defects before production. This policy-driven approach works across OpenAI-compatible stacks as well as frameworks like LangChain, CrewAI, and others, giving platform teams a consistent way to validate agents. Tool governance is handled through Toolboxes, which centralize authentication, lifecycle, and access control for tools and data, while skills stay discoverable but contained by project. Foundry also adds shared observability and direct publishing into Microsoft Teams and Microsoft 365 Copilot, with identity, permissions, and policy applied automatically. For enterprise developers, this means a unified, managed SaaS experience for AI agents production: design, test, deploy, monitor, and enforce enterprise AI governance without stitching together many disconnected services.
