// the find
microsoft/agent-framework
A framework for building, orchestrating and deploying AI agents and multi-agent workflows with support for Python and .NET.
Microsoft Agent Framework is Microsoft's official library for building multi-agent AI systems in Python and .NET. It targets teams moving agents from proof-of-concept to production workloads that need orchestration, durability, and observability. It's positioned as the successor to both Semantic Kernel's agent abstractions and AutoGen.
First-class .NET support is rare in the agent framework space — most competitors are Python-only, so this matters for enterprise teams already on C#. The graph-based workflow model (sequential, concurrent, handoff, group collaboration) with checkpointing and time-travel is a real feature, not just a buzzword — the ADRs in docs/decisions show actual design work behind it. Built-in OpenTelemetry integration means you get distributed traces without bolting on your own instrumentation. The declarative YAML agent format and the DevUI for interactive debugging are practical additions that reduce the iteration loop.
It's deeply Azure-first — the quickstart literally requires `az login` and the path of least resistance routes everything through Microsoft Foundry, Azure OpenAI, and DefaultAzureCredential. Using it with a non-Azure provider is possible but you're immediately swimming against the current. The migration guides from Semantic Kernel and AutoGen exist, which implies Microsoft has already deprecated two previous frameworks in this space — adopting this one carries a real risk of a third rewrite in a few years. The Python and .NET APIs claim to be 'consistent' but dual-language maintenance at this scope historically means one side lags; the ADR list is skewed toward Python decisions, suggesting .NET may be catching up rather than leading. No clear story for self-hosted or air-gapped deployments — everything assumes cloud connectivity.