Microsoft’s agent toolkit roundup

Microsoft is now pitching its agent stack as three layers: Semantic Kernel for app orchestration, AutoGen for multi-agent teamwork, and Agent 365 for enterprise controls around agents inside Microsoft 365. (learn.microsoft.com) An agent is software that can call tools, keep context, and take multi-step actions instead of answering a single prompt. Microsoft’s current documentation says Semantic Kernel handles plugins, memory, and process flows, while AutoGen focuses on agents that cooperate through messages and shared workflows. (learn.microsoft.com ) (microsoft.com) Semantic Kernel’s plugin system is Microsoft’s wrapper around callable functions, and its memory features let developers save and recall information across turns. Microsoft’s Learn pages list plugins, memory, process framework, observability, and security as core building blocks. (learn.microsoft.com 1) (learn.microsoft.com 2) AutoGen came out of Microsoft Research and is built for cases where several agents divide work, debate options, or hand tasks to one another. Microsoft says the framework uses an asynchronous, event-driven architecture with observability, reusable components, and support for Python and.NET. (microsoft.com) Agent 365 sits on a different layer: governance. Microsoft’s product page says it gives IT teams one place to observe, govern, and secure agents across an organization, and Microsoft says the service will be generally available on May 1, 2026. (microsoft.com) The identity model is the part that pushes agents closest to employee-like treatment. Microsoft’s Agent 365 identity documentation says each agent gets a persistent enterprise identity, separate from human users, and those “agentic users” can be synchronized into a tenant, assigned licenses, and given mailboxes or OneDrive storage based on those licenses. (learn.microsoft.com) That licensing detail helps explain why Microsoft’s demos keep drifting from code into procurement. Microsoft’s April 2026 Copilot Studio licensing guide says customers need to match product and usage rights to specific agent scenarios and should use the latest guide or a Microsoft account team for exact requirements. (microsoft.com) The technical story also moved again over the past six months. Microsoft now describes Microsoft Agent Framework as the successor path for agent development, saying it combines AutoGen’s agent abstractions with Semantic Kernel’s state management, middleware, telemetry, and graph-based workflows. (learn.microsoft.com) (devblogs.microsoft.com) That means the roundup lands in the middle of a transition, not at the start of one. Microsoft says Semantic Kernel will keep getting support for existing users, but the newer Agent Framework is its “single call-to-action” for developers building production agents going forward. (devblogs.microsoft.com) The C# angle in Microsoft’s demos fits that strategy. Microsoft’s current Agent Framework documentation includes OpenAI support in C# for Chat Completions, Responses, and Assistants, and its.NET AI quickstarts show how Microsoft is courting developers who want to bolt agent behavior onto existing business apps. (learn.microsoft.com 1) (learn.microsoft.com 2) So the current Microsoft message is less “here is one agent product” than “here is a stack, plus a migration path.” Build the behavior with frameworks, wire it into work apps, then hand the resulting agents to the same identity, security, and admin systems Microsoft already sells for people. (learn.microsoft.com) (microsoft.com)