From Demo-Ware to Durable AI Agents
Microsoft Foundry is a unified Microsoft Foundry platform for building, grounding, and governing AI agents so they can run safely, repeatably, and at scale inside enterprise systems rather than remaining fragile prototypes. The latest Build announcements frame Foundry as the enterprise AI infrastructure layer where AI agents move from experiments to production, with an emphasis on reliability over raw model capability. Instead of shipping another set of model endpoints, Microsoft added what Nick Brady describes as “runtime, tools, memory, grounding, models, observability, and governance” that production teams expect. The focus is clear: AI agents in production must survive real workloads, real data, and real compliance checks. This approach positions Foundry as an operating environment for AI agents in production, where consistent behavior, policy alignment, and observability matter more than one-off benchmark scores.
Hosted Runtime: Agents Without a Rewrite
At the center of the new AI agents production story is Foundry Agent Service, a hosted runtime that gives each agent session its own sandbox, compute, memory, and durable filesystem. Enterprise developers can treat it as shared enterprise AI infrastructure rather than running bespoke stacks per project. A key design choice is interoperability: agents built on Microsoft Agent Framework, GitHub Copilot SDK, LangGraph, and other SDKs can deploy without rewrites. Two protocols target different teams. The Responses API supports OpenAI-style stateful conversations, while the invocations protocol preserves custom request and response formats for existing orchestrations. Routines, now in public preview, add scheduled runs for recurring jobs like nightly ticket triage or daily summaries. According to InfoQ, the same runtime also supports long-running agents with durable state, building on earlier Azure AI Foundry Agent Service work.
Toolboxes, Memory, and the Knowledge Layer
Tool use and knowledge integration often become hidden sources of fragility for AI agents in production. Toolboxes in Foundry, now in public preview, address this by giving each agent a single managed endpoint for tools, skills, Model Context Protocol clients, and enterprise integrations. Teams register tools once, then let agents discover them at runtime with centralized AI agent governance over auth and lifecycle. Skills are versioned and cataloged per project, and a tool search feature narrows the tools exposed to each task, which improves answer quality and keeps context windows lean. Memory moves from app responsibility into the platform: Foundry Agent Service now supports user, session, and procedural memory that helps agents learn how to perform work across runs. Microsoft reports early Tau benchmarks where procedural memory raises task success rates by 7 to 14 percent at near-baseline cost.
Governance, ASSERT, and Logic Apps Automation
Governance is where Microsoft is drawing a sharper line between Foundry and more ad hoc AI stacks. ASSERT, an open-source framework from Microsoft Research, converts written policies into measurable evaluations for agents built with LangChain, CrewAI, LightLLM, OpenAI, and others. Instead of relying on static benchmarks, teams can generate targeted scenarios that expose safety and quality issues before deployment. Toolboxes also plug agents into Microsoft IQ, including Work IQ, Fabric IQ with Fabric’s data agent, Ontology, and semantic models, so agents can access enterprise data through a single, SLA-backed knowledge layer. On the orchestration side, Logic Apps Automation brings workflows, AI agents, knowledge services, and model access into a managed SaaS experience on Azure, making it easier for operations teams to standardize how agents are integrated into business processes while preserving consistent policies across environments.
Why Foundry’s Reliability-First Strategy Matters
For enterprise developers, the main shift is that Microsoft Foundry can now serve as a unified platform instead of a patchwork of bespoke services. Hosted agents, Toolboxes, shared memory, Foundry IQ, ASSERT, and Logic Apps Automation combine into a consistent enterprise AI infrastructure layer where reliability and governance come first. Microsoft appears to bet that the enterprise AI battle will be decided less by whoever has the flashiest model and more by who can keep large fleets of agents operating safely and predictably under real-world constraints. Rather than forcing teams to choose between agility and control, Foundry’s runtime and open protocols let existing agents move into production with minimal rewrites. For organizations that have struggled to graduate promising proof-of-concept agents into live systems, this shift could mark the difference between scattered experiments and a governed, production-grade AI estate.
