Microsoft Build 2026 sharpens Azure’s agent stack

Microsoft Build 2026 centered a major share of its developer agenda on agents, spanning new in-house models, managed runtime infrastructure and supporting services for memory, retrieval, optimization and governance. Microsoft introduced new MAI models trained from scratch on clean and commercially licensed data, covering reasoning, coding, image generation and editing, transcription, and text-to-speech.

MAI-Thinking-1 is positioned as the reasoning flagship, with a 35B active-parameter mixture-of-experts design and a 256K context window. MAI-Code-1-Flash is a smaller agentic coding model rolling out as one of the default models in VS Code, while MAI-Transcribe-1.5 claims transcription across 43 languages and performance up to 5x faster than rival models.

Foundry Agent Service gained hosted agents for running containerized agent code on Azure with managed scaling, observability, sandboxed sessions and per-agent Entra identity. Hosted agents are set to reach general availability in July 2026, alongside public preview features including Routines for scheduled or triggered execution and Toolboxes for governed MCP tool bundles.

Microsoft also expanded Foundry with Agent Optimization, managed memory, Foundry IQ retrieval and AI Gateway controls in Azure API Management. Memory now covers procedural, user and session scopes, with procedural memory described as improving task success 7-14%, while the gateway adds unified model access, routing, fallback, semantic caching and logging for prompts, completions and tool calls.