↓ Skip to main content
Brian Swiger

What's New at Microsoft Build 2026: Building, Deploying, and Operating Enterprise Agents at Scale

·4 mins
Cloud Architecture Artificial Intelligence Microsoft Ecosystem Microsoft Build 2026 AI Agents Microsoft Foundry GitHub Azure AI Enterprise Automation
Brian Swiger
Author
Brian Swiger
Passionate Geek • Proud Father • Devoted Husband
Table of Contents

The transition from experimental AI prototypes to production-grade enterprise assets has historically been plagued by fragmentation. Developers often find themselves juggling separate tools for local authoring, cloud hosting, enterprise distribution, and continuous optimization.

At Microsoft Build 2026 in San Francisco, Microsoft addressed this friction head-on by launching the Microsoft Foundry Agentic Platform. This release delivers a fully connected, three-layer architecture: Build in GitHub, Run in Foundry, and Reach users across Microsoft 365.

1. Build: Unified Frameworks and Smarter Memory
#

The developer experience receives a major upgrade with the stabilization of core frameworks and tighter tooling integration.

  • Microsoft Agent Framework (Stable Release): This is Microsoft’s opinionated, open-source SDK that officially unifies Semantic Kernel and AutoGen. It includes a robust agent harness equipped with skills and memory, native integrations with the GitHub Copilot SDK and Claude Agent SDK, and advanced multi-agent orchestration patterns like Magentic-One.
  • Foundry Toolkit for VS Code (GA): Developers can now build, debug, inspect, and deploy agents directly to the Foundry Agent Service without leaving their primary IDE.
  • Toolboxes in Foundry (GA Soon): Provides a single managed Model Context Protocol (MCP) endpoint for all tool types. Skills are now versioned and discoverable, featuring intelligent tool search to dynamically select the right utility for a given task. Grounding data is supported via native connections to Work IQ, Fabric IQ, Web IQ, and Foundry IQ.
  • Advanced Foundry Memory (Public Preview): To combat context loss, the platform introduces procedural memory, alongside user and session memory. Early Tau-bench evaluations demonstrate a strict 7% to 14% absolute success-rate gain in agent task execution.

2. Deploy: Managed Runtimes and Org-Chart Integration
#

Deploying an agent safely inside an enterprise requires strict boundary isolation and deep identity management.

Foundry Agent Service now offers Hosted Agents (GA soon), providing a fully managed, session-isolated runtime. The runtime is completely framework-agnostic. Whether an agent is built using the Microsoft Agent Framework, GitHub Copilot SDK, or LangGraph, it can be hosted without code rewrites. It also introduces public preview support for long-running autonomous agents and scheduled routines.

Reaching the Workspace
#

Distribution is split into two primary paradigms:

  1. App Integration (GA Soon): One-click publishing to Microsoft Teams and Microsoft 365 Copilot, where organizational identity, permissions, and compliance policies flow through automatically.
  2. Autopilot Agents (Public Preview): This represents a new tier of digital worker. These agents are provisioned with an Entra Agent ID, an enterprise email address, a functional Teams presence, and a literal placement within the corporate org chart. Governed by Agent 365, they can collaborate independently with humans and other systems.

Furthermore, the Incoming A2A (Agent-to-Agent) public preview allows any Foundry agent to be exposed as an endpoint. Other agents can discover it via its digital “agent card” and invoke it using an open A2A protocol, cutting across different frameworks and cloud boundaries.

3. Operate: Enterprise Observability and Closed-Loop Optimization
#

Operating agents at scale requires the same rigor as traditional microservices, demanding deep tracing and continuous optimization.

Tracing and Evaluation
#

The platform introduces an end-to-end OpenTelemetry pipeline for hosted agents (GA soon). Every LLM call, tool invocation, and sub-agent hop is logged. These evaluations link directly back to the exact trace that produced them inside the Foundry Control Plane.

The Agent Optimizer
#

Entering public preview within the next 30 days, the Agent Optimizer automates the post-deployment refinement cycle. The system ingests production traces and evaluations, generates ranked candidate improvements for prompts and skills, tests those candidates against historical scenarios, and recommends changes with a full audit log and rollback lineage.

[ Production Traces & Evals ] 
             │
             ▼
     [ Agent Optimizer ] ───► Generates & Ranks Improvements
             │
             ▼
   [ Scenario Validation ] ──► Full Lineage, Diff, & Audit
             │
             ▼
   [ Recommended Deploy ]

Architecture and Strategy Architecture
#

This release establishes Microsoft Foundry as a highly integrated platform covering the entire agent lifecycle. While market alternatives offer point solutions for orchestration or hosting, this ecosystem unifies local design, enterprise-scale runtime execution, native Microsoft 365 distribution, and automated optimization under a singular governance framework via Agent 365.

Official Resources and Documentation
#

To explore the implementation details, review the documentation provided by Microsoft engineering: