Microsoft has fundamentally reshaped how enterprises deploy AI agents in production, moving beyond experimental prototyping into reliable, secure, real-world systems. The company's Foundry Agent Service is now generally available, built on OpenAI's Responses API and compatible with models from DeepSeek, xAI, Meta, and others. This shift matters because it solves a critical gap: most organizations can build AI agents in a lab, but deploying them safely at scale remains a puzzle .

What Changed in Microsoft's Agent Framework?

The Foundry Agent Service launches with several production-grade features that address real deployment challenges. The runtime is built on the Responses API, meaning developers using OpenAI agents can migrate to Foundry with minimal code changes. What they gain in return is enterprise-grade infrastructure: end-to-end private networking, Entra role-based access control (RBAC), full execution tracing, and continuous quality monitoring piped directly into Azure Monitor .

The architecture supports hosted agents across six new regions, giving teams geographic flexibility for latency-sensitive workloads. Voice Live, a fully managed speech-to-speech runtime, collapses the traditional pipeline of speech-to-text, language model processing, and text-to-speech into a single managed API. This includes semantic voice activity detection, end-of-turn detection, server-side noise suppression, and barge-in support, all built-in .

How to Deploy AI Agents Securely in Production

• Governed Tool Registry: Register Python functions as tools through a central, governed registry rather than giving language models raw access to your systems. This ensures agents only access data they have explicit permissions to use.
• Observability and Tracing: Capture detailed operations, inputs, and timing data through execution spans to gain absolute visibility into the agent's decision-making path and tool selection when mistakes occur.
• Predictable API Interfaces: Define production agents with predictable, versioned APIs rather than ad-hoc prompts, ensuring consistent behavior across deployments and easier debugging.
• Continuous Monitoring: Use out-of-the-box and custom evaluators with continuous production monitoring to track quality as a live signal, not just a pre-deployment checkbox.
• Third-Party Security Integration: Layer in runtime security tools like Palo Alto Prisma AIRS and Zenity for prompt injection detection, data leakage prevention, and tool misuse detection .

Why Model Pricing and Sizing Matter for Agent Economics

Microsoft released three new model tiers in March 2026, each designed for different agent workloads. GPT-5.4, which went generally available on March 5, is specifically engineered for production reliability rather than raw intelligence. The focus is on solving real problems: task drift, mid-workflow failures, and inconsistent tool calling. Stronger reasoning over long interactions, better instruction adherence, and integrated computer use capabilities enable structured orchestration of tools, files, and data extraction .

Pricing reflects this specialization. Standard GPT-5.4 costs $2.50 per million input tokens and $15 per million output tokens, with a cached input tier at $0.25 per million tokens. For larger contexts exceeding 272,000 input tokens, pricing extends to $5.00 per million input tokens and $22.50 per million output tokens. GPT-5.4 Pro, the premium variant for analytical depth, costs $30 per million input tokens and $180 per million output tokens, targeting scientific research and complex trade-off analysis where reasoning quality matters more than latency .

GPT-5.4 Mini, released March 17, offers a cost-efficient tier for high-volume, low-latency tasks like classification, extraction, and lightweight tool calls. This enables a routing strategy where Mini handles the high-volume tier while full GPT-5.4 takes reasoning-heavy work, optimizing both cost and performance .

What About Open-Source and Alternative Models?

Microsoft's Foundry now integrates open-source models directly into the same deployment and management flow as proprietary models, eliminating context-switching between providers. Fireworks AI brings high-performance open model inference to the platform, processing over 13 trillion tokens daily at approximately 180,000 requests per second in production. Four models are available at launch: DeepSeek V3.2 with sparse attention and 128,000-token context, OpenAI's open-source gpt-oss-120b, Moonshot AI's Kimi K2.5, and MiniMax M2.5 .

The real innovation here is bring-your-own-weights support. Teams can upload and register quantized or fine-tuned weights from anywhere without changing the serving stack, deploying via serverless pay-per-token or provisioned throughput. NVIDIA Nemotron models, announced at NVIDIA GTC on March 16, are now first-class citizens in the Foundry catalog, available on NVIDIA accelerators. Combined with Fireworks AI integration, teams can fine-tune Nemotron into low-latency assets distributable to the edge .

Why Enterprises Are Moving Toward Agent-First Development

The broader context matters here. Uber recently demonstrated how AI-powered prototyping can compress a month of cross-functional debate into a two-hour workflow, generating clickable flows and interactive, localized demos in hours rather than weeks. This represents a fundamental shift in how enterprise tech companies skip traditional bottlenecks to achieve instant internal alignment .

However, as organizations move toward an agent-first economy, security becomes paramount. New research into "AI Agent Traps" reveals that securing the model alone is insufficient; teams must also secure the environment. From semantic manipulation to knowledge base poisoning in retrieval-augmented generation (RAG) systems, the data that agents "read" is being weaponized. This is why Microsoft's emphasis on tracing, evaluation, and third-party security integrations represents a maturation of the field .

The SDK consolidation also matters for developer experience. The azure-ai-projects SDK reached version 2.0.0 across Python, JavaScript, TypeScript, and Java, with .NET 2.0.0 following on April 1. The azure-ai-agents dependency is gone; everything now lives under AIProjectClient. Developers simply run "pip install azure-ai-projects" and the package bundles OpenAI and Azure identity as direct dependencies .

For teams already using the Responses API with OpenAI agents, migration is straightforward. The code example shows minimal changes: define an agent with a prompt and model, create a conversation, and pass the agent reference in the API call. This compatibility reduces friction for enterprises considering a move to Microsoft's infrastructure .

The deprecation of PromptFlow, with migration to Microsoft Framework Workflows required by January 2027, signals Microsoft's commitment to consolidating its agent development stack. Combined with continuous monitoring, tracing, and evaluation now in general availability, the message is clear: AI agents are no longer experimental. They are production infrastructure, and enterprises need the tooling to match .