From Impressive Demos to Reliable Enterprise AI Agents
Microsoft Foundry is Microsoft’s AI app and agent factory that focuses on turning experimental enterprise AI agents into reliable, governed systems ready for AI production deployment across complex organizations, providing shared runtime, memory, grounding, tooling, and policy controls so teams no longer have to assemble fragmented infrastructure on their own. The agentic wave has produced many demos, but fewer agents that survive real load, real data, and real compliance demands. At Build, Microsoft shifted Foundry from a model-first preview into something closer to an enterprise runtime. Instead of racing on capability alone, the Microsoft Foundry platform now tries to differentiate on infrastructure stability, observability, and AI governance tooling. This reflects how buyers are changing: they care less about the newest benchmark score and more about whether an agent can run every day, pass audits, and integrate with existing tools and identity systems.
Hosted Runtime: Sandboxed Sessions Without Rewrites
A cornerstone of Foundry’s agent reliability infrastructure is the hosted runtime in Foundry Agent Service, which moves agents from ad‑hoc hosting to managed, sandboxed sessions with state and filesystem access. Each agent session runs with its own dedicated compute, memory, and durable storage, and long‑running agents keep state across runs for workloads like OpenClaw or Hermes. Crucially, the runtime is framework‑agnostic: agents built with Microsoft Agent Framework, GitHub Copilot SDK, LangGraph, and other SDKs can be deployed without rewrites, using either a stateful Responses API or a pass‑through invocations protocol. That design respects teams that already have orchestration in place. Routines, now in public preview, add scheduled execution for overnight ticket triage or daily reporting, making AI production deployment feel closer to traditional job scheduling. Together these pieces turn Foundry from a model endpoint collection into a general agent runtime.
Toolboxes and Memory: Platform-Level Capabilities, Not One-Off Plumbing
Foundry shifts key agent capabilities into the platform so developers stop rebuilding the same patterns. Toolboxes, now in public preview, give each agent a single managed endpoint for tools, skills, Model Context Protocol clients, and enterprise data integrations. Teams register tools once; at runtime, agents discover them with central auth, lifecycle, and governance. Skills become versioned, project-scoped artifacts, and tool search aims to select a focused tool set per task instead of dumping an entire catalog into a context window. On the memory side, Foundry Agent Service exposes procedural, user, and session memory as shared services. Procedural memory helps agents learn how to perform 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 procedural memory is enabled, directly tying platform features to measurable task reliability.
Governance and Knowledge: From Benchmarks to Policy-Driven Control
Governance in the Microsoft Foundry platform moves beyond static benchmarks toward policy‑driven control. ASSERT, an open-source framework for evaluation and regression testing, converts written policies into concrete, measurable checks, generating targeted scenarios to catch safety and quality defects before they reach production. It works across frameworks including LangChain, CrewAI, LightLLM, and OpenAI, so teams can standardize evaluation even in mixed stacks. Governance also reaches into tools through Toolboxes, where skills and MCP resources are cataloged, versioned, and discoverable under consistent rules. For knowledge and grounding, Foundry IQ acts as a unified layer behind agents, connecting Work IQ, Fabric IQ, Azure SQL, file search, and other sources under a single SLA-backed retrieval endpoint. This combination of policy-aware evaluation and centrally managed knowledge bases pushes Foundry toward being a control plane for enterprise AI agents rather than just another place to call models.
A Maturing Market: Reliability as the Competitive Battleground
The latest Foundry release signals a broader market shift: enterprises now treat AI production deployment as an engineering discipline, not a lab experiment. Observability, policy, identity, and scheduled workloads are table stakes; model choice is one input, not the main stage. Foundry’s promise is to become the infrastructure layer enterprises previously stitched together piecemeal: sandboxed runtimes, shared memory, unified data access, AI governance tooling, and distribution into Microsoft Teams and Microsoft 365 Copilot with identity and policy applied automatically. By framing Foundry as “the place where AI agents move from experiments to production systems,” Microsoft is betting that the next competitive front in enterprise AI agents is reliability and governability rather than raw model capability. If that bet holds, success stories will be measured less by eye‑catching demos and more by quiet agents that run every day, comply with policy, and keep failing less over time.
