AI agents on LangGraph.

LangGraph gives you an explicit state machine - nodes, edges, conditional transitions, interrupt points. CrewAI and AutoGen abstract that into role-based or message-passing models, which feel cleaner in a demo and harder to reason about under load. For workflows with branching, human-in-the-loop, or strict latency budgets, the explicit graph is worth the verbosity. For one-off multi-agent demos, the abstractions win.

Two layers. (1) Versioned state schemas - every checkpoint carries the schema version it was written under. (2) A migration step that reads old-version checkpoints, transforms them, and writes them back at the current schema version. We design this before any state touches durable storage, because retrofitting it under production traffic is brutal.

OpenAI Agents SDK wins if you want low ceremony and are comfortable on the OpenAI stack - clean API, built-in tracing, minimal boilerplate, ships fast. LangGraph wins when you need explicit state control: a persisted, versionable state graph; conditional branching that's legible in code; human-in-the-loop interrupts at specific graph nodes; and per-node observability that shows exactly which step failed and what state it held. The practical gap appears in your first production incident. On LangGraph, a bad run is a trace of the exact failing node plus its state. On the Agents SDK, you inspect the thread and guess. If your agent is multi-step, stateful with persistent conversations, or has strict latency budgets per branch - LangGraph's verbosity pays back at the first incident. For one-off tools or simple linear flows, the Agents SDK is the faster path.

MCP servers wrap cleanly as LangChain tools - a BaseTool subclass with an MCP client in the run method. The graph node that calls it gets the same retry logic, error handling, and OTel tracing as any other tool. The piece teams consistently miss: MCP servers with persistent session state (SSE connections, session IDs) need explicit lifecycle management across a multi-worker LangGraph deployment. Each worker needs its own managed client with connection pooling and health checks, or you'll hit session collisions under concurrent load. We wire this alongside the Postgres checkpointer design in week one - because adding it later, under live traffic, is a painful retrofit.