Microsoft Agent Framework

What people say

Agent Framework 1.0 shipped April 7, 2026, closing out roughly a year of Microsoft running two separate, half-competing agent SDKs. The pitch: AutoGen's simple agent abstractions plus Semantic Kernel's enterprise features — type safety, middleware, telemetry, session-based state — in one supported framework, with orchestration patterns (sequential, concurrent, handoff, group chat, Magentic-One) that all support streaming, checkpointing, and pause/resume for long-running workflows.

Early developer writeups single out the graph-based workflow engine as the standout addition. One hands-on account, written three days into using it, described defining agent relationships as a directed graph — each agent keeping its own session state — as a cleaner mental model than AutoGen's conversation threads. The local DevUI debugger drew similar praise for surfacing real-time message flow between agents, useful enough that it caught an infinite-loop bug that would otherwise have run up API costs unnoticed.

Migration guidance from Microsoft's own devblog puts Semantic Kernel migration at 2-4 hours for a typical app. AutoGen migration is harder, since the conversation-centric model doesn't map cleanly onto the new graph structure — teams with heavy AutoGen investment should budget more time than the marketing implies.

The framework isn't without security scrutiny. A Hacker News-covered flaw in AutoGen Studio (the prototyping UI, not Agent Framework itself) let a webpage hijack an agent for local code execution via an unauthenticated MCP WebSocket handler — the "AutoJack" attack. Microsoft's security team traced the root cause to a broader pattern across agent frameworks: treating localhost as a trust boundary when an agent that browses the open web effectively erases that boundary. The specific handler has since been hardened, but it's a reminder that the surrounding agent tooling carries its own risk surface independent of the core SDK.

Cost-wise, the framework itself is free and open source under MIT. Running agents through Microsoft Foundry / Azure AI Foundry incurs standard token and compute charges — the platform layer is free to explore, but model consumption, Foundry Tools connectors (Logic Apps, SharePoint, Bing grounding) and Foundry IQ data connections bill separately. For teams already inside Azure's ecosystem, that's a known-quantity cost model rather than a surprise.

Summary of public user & expert reviews, compiled by RECATOOLS.