From Playground Bots to Production AI Agents
Microsoft Foundry agents are long‑running, multi‑step AI programs that run on a managed platform, connect to enterprise data and tools, and are monitored, governed, and deployed like any other critical production service rather than as ad‑hoc prompt experiments. Microsoft has moved this idea from vision to reality by making production agents in Foundry Agent Service generally available, including hosted agents, toolboxes, and publishing into Microsoft 365 Copilot and Microsoft Teams. More than 100,000 organizations are already building on Microsoft Foundry, and companies such as Adobe, Telefónica, and Tata Consultancy Services are running agents in production today. This is not a pilot; it is a signal that AI is now expected to live inside reliable systems, with identity, security, and compliance wired in from day one.
Frontier Models and Data Zones: The New Deployment Surface
The most important part of this launch is not the feature checklist but what it says about where enterprise automation is heading. OpenAI’s GPT‑5.6 Sol, Terra, and Luna models are now generally available inside Microsoft Foundry Models and Agent Service, giving teams a spectrum from advanced reasoning to high‑volume, low‑latency work on one platform. GPT‑5.6 Sol targets extended reasoning and agentic workflows; Terra balances performance and cost for everyday work; Luna prioritizes speed and affordability for latency‑sensitive tasks. Pricing reflects this spread, from USD 5.00 (approx. RM23) per million input tokens for Sol to USD 1.00 (approx. RM4.60) for Luna, with output tokens priced up to USD 30.00 (approx. RM138) for Sol and USD 6.00 (approx. RM27.60) for Luna. At the same time, the new Asia‑Pacific Data Zone lets organizations run these frontier models while keeping processing in‑region, aligning AI adoption with sovereignty and compliance requirements without fragmenting their development experience.
Long‑Running Workflows: Agents That Don’t Forget
The real shift is that production AI agents are now expected to survive time, failure, and organizational complexity. Hosted agents in Microsoft Foundry are generally available and provide “agent‑native compute” for any framework or language, giving teams a single production runtime for Microsoft Agent Framework, GitHub Copilot SDK, and others. For long‑running workloads, new resilient task support in hosted agents makes it easier to build agents that outlast process crashes and infrastructure hiccups, keeping multi‑turn conversations, reasoning loops, and human‑in‑the‑loop approvals intact without custom recovery logic. Routines allow these agents to run on schedules or timers so they can act without a user prompt, and event‑based triggers can wake them the moment a ticket is filed, a file lands in storage, or an upstream workflow completes. This is what long‑running workflows look like when they stop being demos and become real automation: persistent state, controlled triggers, and observable behavior.
GPT‑5.6 with Work IQ: Reasoning Meets Context
GPT‑5.6 with Work IQ threads stronger reasoning directly into the tools where people already work. GPT‑5.6 Sol brings extended reasoning and agentic workflows to demanding workloads, while Terra and Luna cover everyday and latency‑sensitive tasks. According to Microsoft, “GPT‑5.6 Terra is a balanced model for everyday work, delivering performance competitive with GPT‑5.5 at a lower cost.” Work IQ is part of Microsoft IQ, which unifies Work IQ for real‑time awareness of Microsoft 365, Fabric IQ for structured data, and Web IQ for live web grounding into Foundry IQ, now the SLA‑backed knowledge layer for every Foundry agent. In practical terms, this means Copilot Chat, GitHub experiences, Microsoft 365 apps, and Foundry agents can all draw from the same governed context, while publishing those agents into Microsoft 365 Copilot and Teams becomes generally available next week.
Beyond Prompts: AI Governance Infrastructure as the Real Product
The most opinionated takeaway from this launch is that prompt engineering is no longer the main event; production AI is a systems problem. Recent industry discussions argue that the hard problems have shifted to context, data contracts, LLM and MCP gateways, state management, evaluation, latency, cost, observability, security, and trust. Microsoft Foundry leans into this by making governance a first‑class pillar—build, generate, govern—rather than leaving it to ad‑hoc scripts. The platform frames trust as a platform priority, insisting that “an agent you can’t see, improve, or secure is an agent you can’t put in production.” That matches the move toward agent harnesses and control planes where teams own state, order mutations, and prove actions, not just prompts. Memory, routines, and an agent optimizer in Foundry Agent Service are already in public preview, pointing toward a future where AI governance infrastructure is as standard as CI/CD pipelines.
