From impressive demos to dependable production AI agents
Microsoft Foundry is an enterprise AI platform that supplies a managed runtime, shared tools, and policy-driven governance so AI agents can move from experimental prototypes and isolated demos into reliable, auditable production systems that work with real business data, workloads, and compliance requirements at scale. At Build, Microsoft framed the enterprise AI battle as one of reliability and governance instead of raw model capability, positioning the Microsoft Foundry platform as an “AI app and agent factory” rather than another set of endpoints. Nick Brady, senior program manager for developer experience, described this release as bringing “runtime, tools, memory, grounding, models, observability, and governance” that teams need once pilots have to support real users. The platform’s goal is to replace the ad-hoc infrastructure many enterprises have been stitching together with a consistent layer for building, deploying, and controlling production AI agents.
Hosted runtime: a production home for diverse AI agent frameworks
The new Foundry Agent Service gives enterprises a managed runtime where each AI agent session runs in its own sandbox with dedicated compute, memory, and durable filesystem access. Hosted agents support frameworks including Microsoft Agent Framework, GitHub Copilot SDK, LangGraph, and others without forcing rewrites, addressing a common migration barrier from prototypes to production AI agents. Two protocols cover most deployment patterns: a stateful Responses API compatible with OpenAI-style interactions, and an invocations protocol for pass-through calls where teams keep full control over request and response formats. Long-running agents can keep durable state and files, while routines, now in public preview, let agents run on schedules for tasks such as overnight ticket triage or daily reporting. Together, these features turn what were often brittle, custom runtimes into a standardised environment for AI agent deployment and lifecycle management.
Toolboxes and memory: taming tools, skills, and enterprise context
As AI agents gain more tools, unmanaged sprawl becomes a reliability and governance problem. Toolboxes in the Microsoft Foundry platform introduce a single managed endpoint for tools, skills, Model Context Protocol clients, and enterprise data integrations, so teams configure once and agents discover at runtime instead of wiring every tool into every agent. Skills are versioned in project-scoped catalogs and surfaced as MCP resources, while tool search helps limit each task to a focused set of tools, which improves response quality and keeps prompts from bloating. Foundry also treats memory as a core service, not an application afterthought. Procedural, user, and session memory live in the Agent Service, with procedural memory teaching agents how to complete work across runs. According to Microsoft’s Nick Brady, early Tau benchmarks show procedural memory delivering “7 to 14 percent absolute success rate gains at near baseline cost.”
Foundry IQ and distribution: grounding agents in real business knowledge
To avoid hallucinations and brittle integrations, Microsoft Foundry uses Foundry IQ as a shared knowledge layer behind AI agents. Foundry IQ unifies Work IQ, Fabric IQ, Azure SQL, file search, and other sources behind a single retrieval endpoint, so agents do not need custom connectors for each data store. At Build, Microsoft announced Foundry IQ Serverless in public preview, multi-source knowledge bases in general availability, and Microsoft Web IQ for live web information, all covered by platform-level service expectations. Toolboxes connect to Microsoft IQ services so agents can query enterprise data, ontologies, and semantic models through governed, discoverable interfaces. On the distribution side, Foundry adds direct publishing into Microsoft Teams and Microsoft 365 Copilot, with general availability planned for June 2026, letting production AI agents appear where employees already work while inheriting identity, permissions, and policy controls automatically.
ASSERT and control specs: enterprise AI governance by design
Enterprise AI governance is where Microsoft is trying to differentiate Foundry. Instead of only citing static benchmarks, ASSERT (Adaptive Spec-driven Scoring for Evaluation and Regression Testing) converts written policies into measurable evaluations for AI agents. Built on Microsoft Research work and released as an open-source framework, ASSERT generates targeted scenarios to expose safety and quality issues before agents reach production. It works across common frameworks such as LangChain, CrewAI, LightLLM, and OpenAI, helping enterprises evaluate behaviour against their own rules. Alongside ASSERT, Microsoft is promoting an Agent Control Spec approach that lets teams declare what agents are allowed to do, which tools they may call, and how outputs should be constrained. Combined with shared observability, these governance tools turn Foundry into an infrastructure layer designed for regulated industries and mission-critical deployments, where auditability and control are as important as capability.
