LangGraph's own team says stop graph-shaping open-ended work — GPT Researcher moved its multi-agent pipeline off graphs
A July 22 retrospective from Sydney Runkle and Harrison Chase draws a usable decision line: use a graph when you can encode real world-knowledge about stage order (their example is classify → search → synthesize), and don't when planning and delegation should emerge at runtime — citing GPT Researcher swapping its graph-shaped multi-agent deep-research pipeline for Deep Agents. Two concrete carry-overs: production agent graphs are not DAGs (retries, human input, validation loops are cycles, so design for them), and the `Send` API is the escape hatch for map-reduce where worker count is only known at runtime. The framing shift is that a node can now hold a full agent run, so you layer fixed steps, model steps, and agent steps in one system rather than choosing.
↳ Follow the thread