A detailed analysis of how to think about AI agent frameworks, written by LangGraph's creator in response to OpenAI's guide on building agents. Covers the distinction between workflows and agents, declarative vs imperative approaches, and why most production agentic systems combine both. Argues that the core challenge in building reliable agents is ensuring the LLM receives the right context at each step, and that most agent frameworks (including OpenAI's Agents SDK) are merely agent abstractions rather than true orchestration frameworks. Positions LangGraph as a low-level orchestration framework with both declarative and imperative APIs, agent abstractions built on top, and production features like persistence, human-in-the-loop, streaming, and fault tolerance. Also includes a comparison spreadsheet of major agent frameworks.
