Microsoft Agent Framework vs LangGraph

Lineage and design intent

Microsoft's documentation is direct about where Agent Framework comes from: it "combines AutoGen's simple agent abstractions with Semantic Kernel's enterprise features" such as session-based state, type safety, middleware and telemetry, and adds graph-based workflows for explicit multi-agent orchestration. The 1.0 announcement lists single agents and service connectors, middleware, memory and context providers, the workflow engine, multi-agent orchestration patterns, declarative YAML and MCP support as generally available. DevUI, Foundry hosted agents, skills and an opinionated "Agent Harness" were still in preview at that point.

LangGraph starts from the other end. Its documentation calls it a low-level orchestration framework that lets developers mix deterministic, hand-coded steps with LLM-driven steps, and it deliberately ships few opinions about prompts or agent architecture. Higher-level conveniences sit in separate packages: LangChain agents provide prebuilt architectures, and Deep Agents is described as an agent harness built on LangGraph. Teams that want batteries included will find more of them inside Agent Framework itself; teams that want a small core will find LangGraph closer to that.

Durable execution and state

Both frameworks checkpoint at the boundary of a parallel step. Agent Framework creates a checkpoint at the end of each superstep, capturing executor state, pending messages, pending requests and responses, and shared state. Since Python 1.13.0 it also records entry checkpoints so that a whole run is replayable. Resuming into a rebuilt workflow requires the same topology and stable executor identities, a constraint Microsoft documents in detail for .NET agents.

LangGraph persists a snapshot of graph state per thread through a checkpointer, which its docs tie to conversation continuity, human-in-the-loop, time travel and fault tolerance. The storage menus differ in a telling way: Agent Framework's documented Python options are in-memory, local file and Azure Cosmos DB, while LangGraph's are in-memory, SQLite and Postgres. Both allow custom stores, but the defaults show which infrastructure each vendor expects. Microsoft also warns that its file and Cosmos stores use restricted pickle deserialisation and that checkpoint storage must be treated as a trust boundary.

Multi-agent patterns

Agent Framework inherits AutoGen's catalogue. Sequential, concurrent, handoff, group chat and Magentic-One orchestrations are named, stable builders, and the checkpoint documentation shows each with a default workflow name. Workflows can be wrapped as agents and nested.

LangGraph expresses the same topologies as graphs that the developer assembles. That is more work for a standard supervisor pattern and less constraining for an unusual one. Neither project's own documentation offers comparative performance evidence for these patterns, and whether adding agents improves results at all is contested (see multi-agent systems), so pattern availability should not be read as a quality claim.

Ecosystem and deployment

Agent Framework's provider list at 1.0 covered Microsoft Foundry, Azure OpenAI, OpenAI, Anthropic Claude, Amazon Bedrock, Google Gemini and Ollama, with MCP support stable and A2A 1.0 support described as "coming soon"; later .NET release notes refer to A2A agent run modes. Hosting leans on Azure, and the October 2026 Python release added native Foundry hosting.

LangGraph's production story runs through LangSmith. LangSmith Deployment is described as a workflow orchestration runtime for agent workloads with durable execution, streaming and horizontal scaling, offered as managed cloud (Plus plan or above), hybrid, self-hosted with a control plane (Enterprise plan) or a standalone server in Docker or Kubernetes. Its documentation says it also runs agents built with other frameworks. LangGraph names Klarna, Uber and J.P. Morgan as users; adoption figures for Agent Framework beyond repository statistics are not publicly documented in the sources reviewed.