Azure AI Studio became Azure AI Foundry in November 2024. By early 2026, Microsoft’s communications primarily refer to it as “Microsoft Foundry” — a unified platform combining a portal, agent runtime, and model catalogue. The February 2026 update elevated the platform from an experimentation tool to a production-capable solution, introducing multi-agent orchestration, Model Context Protocol support, hosted agents, and sovereign local deployment options.
Key February 2026 Updates
Model Context Protocol Integration
The addition of MCP enables agents to connect to external tools and data sources through standardized interfaces. Organizations can build custom MCP servers linking to inventory databases, CRM systems, and document repositories without requiring custom integration code for each system.
Hosted Agent Infrastructure
Developers can define agents in YAML format, execute two CLI commands, and have Foundry provision compute resources, register endpoints, and provide a URL for production use.
Agent-to-Agent Communication
Multiple specialized agents can coordinate through defined interfaces. Document intake, business rule validation, and system updates can be handled by separate agents rather than one monolithic system, mirroring microservices architecture patterns.
Foundry Local Deployment
Large multimodal models (text, image, audio) now run on local NVIDIA hardware without cloud connectivity, addressing data sovereignty and air-gapped environment requirements.
Resource Provisioning Patterns
Hosted Agents (Managed)
The simplest approach creates: a resource group, Foundry account and project, model deployments (with declared SKU and capacity), Azure Container Registry, and managed identity. A concrete example shows GPT-4o-mini deployment as GlobalStandard with capacity: 10.
Get exclusive access to all things tech-savvy, and be the first to receive
the latest updates directly in your inbox.
Limitation: Hosted agents currently operate exclusively in North Central US as of March 2026.
Bring-Your-Own-Runtime
Organizations deploy conventional App Service, Container App, or AKS workloads while calling Foundry’s /openai/v1/ endpoints directly (GA as of February 2026 for chat completions, embeddings, files, fine-tuning, and vector stores). This pattern separates application runtime from Foundry infrastructure, with AI Search, storage, and Key Vault as supporting services.
Foundry Local
On-premises deployment with zero cloud connectivity. The February 2026 update expanded capabilities from small language models to large multimodal models on NVIDIA GPU hardware.
Deployment Workflow
The azd CLI extension provides a stepwise process:
azd init -t Azure-Samples/azd-starter-basic --location northcentralus
azd ai agent init -m
azd upThe agent.yaml file contains: metadata (name, description, version), implementation templates, environment variables, and user_config specifying the model. The azure.yaml file defines the deployable stack, including Bicep infrastructure, service definitions with container resource settings, and model deployment bindings.
Day-two operations use these commands:
azd ai agent invoke --message "..."azd ai agent runazd ai agent monitor --followazd ai agent showazd down
CI/CD integration uses the Azure/setup-azd GitHub Action, with azd pipeline config generating workflow files and configuring secrets.
Cost Analysis
Token Pricing (Azure OpenAI, Global Standard, Early 2026)
- GPT-4o: $2.50 per million input tokens, $10 per million output tokens
- GPT-4o-mini: $0.15 per million input tokens, $0.60 per million output tokens
The output-to-input cost ratio is 4:1 for both models. Output tokens often represent the larger expense for agents generating structured summaries and action plans.
Cost Tiers
Simple FAQ Chatbot (2,000 chats/day, GPT-4o-mini, small knowledge base)
Model inference: ~$20/month. With AI Search, storage, ingestion pipelines, and monitoring: realistic total $20–$200/month.
Document Processing Agent (500 documents/day)
Azure Document Intelligence costs approximately $10 per 1,000 pages. At 500 two-page documents daily, extraction alone runs ~$300/month. Total system cost: low hundreds to low thousands monthly.
Multi-Agent Workflow Automation (tool-heavy, approvals, long-running orchestration)
February 2026 introduced Durable Agent Orchestration pairing Durable Functions with Agent Framework and SignalR. Conservative range: $1,000–$10,000/month, with higher costs from heavy retrieval, enterprise monitoring, and premium licensing.
Cost Optimization Note: 30–60% of production traffic can often use cheaper models after measurement. Organizations frequently find GPT-4o-mini handles classification and extraction at a fraction of GPT-4o’s cost with equivalent results.
Business Applications
Document Processing
Azure Document Intelligence achieves 96% accuracy on printed text (DeltOCR Bench, January 2026) and 85–86% on invoice processing out of the box. Combined with LLM reasoning and business rule validation, accuracy reaches 97%+. Acentra Health’s clinical correspondence system on Azure OpenAI reported “11,000 nursing hours saved, roughly $800,000 in cost reductions, and document accuracy above 99%.”
Knowledge-Grounded Customer Service
Pairing Azure AI Search (indexing documentation, policies, support history) with OpenAI models grounds responses in indexed information. Well-implemented systems handle 60–80% of routine queries without human intervention.
Multi-Step Workflow Automation
Chaining specialized agents for business-specific processes offers genuine value. Document intake, verification, system updates, and confirmations represent unique business operations that generic SaaS products cannot replicate. Red Eagle Tech, a UK-based consultancy, specializes in this pattern.
Important Constraints
Region Availability
Hosted agents are limited to North Central US as of March 2026. UK and EU businesses face latency and data residency considerations, requiring either acceptance of latency, the bring-your-own-runtime pattern with Foundry endpoints in closer regions, or waiting for expanded availability.
SDK Breaking Changes
The Microsoft Agent Framework reached 1.0.0rc1 in February 2026 with significant renames across credentials, sessions, and response patterns. Earlier beta code will break without migration.
Preview Churn
The azd agent extension remains in public preview. CLI flags, templates, and resource providers may change between updates. Microsoft recommends using GA REST surfaces directly when SDKs lag behind GA releases.
Observability Limitations
Streaming logs via azd ai agent monitor and VS Code’s AI Toolkit Agent Inspector function for local debugging. However, distributed tracing across agent-to-tool-to-retrieval calls is not fully integrated with Azure Monitor out of the box. Organizations must instrument tool calls and correlation IDs independently for multi-agent workflows.
AI Search Virtual Network Constraint
The Azure AI Search tool integration does not support private endpoint connections for services behind a virtual network, requiring workarounds for locked-down VNet expectations.
Implementation Recommendations
- Select a focused process: document intake, customer FAQ handling, or extraction from a specific document type.
- Build a prototype in Agent Builder UI: Connect data, test with real queries, measure output quality at no cost beyond existing Azure subscription.
- Establish baseline metrics: Time, error rate, cost of manual processes before automation enable accurate ROI calculation.
- Deploy incrementally: Agents handle straightforward cases; humans handle exceptions. Expand scope as evaluation metrics support it.
- Treat agents as production software: Evaluate, monitor, version, and document the same way as any engineering system.
References
- Microsoft Foundry February 2026 update — Microsoft Developer Blog
- Azure Developer CLI Foundry agent extension announcement — Azure SDK Blog
- Microsoft Learn: azd AI Foundry extension deployment guide
- Microsoft Learn: Azure AI Search tool integration for Foundry agents
- Azure OpenAI pricing and cost optimization strategies — Finout
- Azure OpenAI Service pricing — Microsoft Azure
- Azure Document Intelligence model overview — Microsoft Learn
- OCR accuracy benchmarks including DeltOCR Bench — AIMultiple
- Intelligent document processing benchmarks — BusinessWareTech
- Acentra Health customer story: AI-powered clinical correspondence — Microsoft
- Power Automate pricing — Microsoft
- Azure Functions Consumption tier pricing — Microsoft Azure
Author: Ihor Havrysh, Software Engineer at Red Eagle Tech, a UK-based firm providing custom AI solutions and bespoke software development on Microsoft Azure.
