LangGraph

I reach for LangGraph when a workflow deserves to be a graph you can read instead of a prompt you can only hope about. Two shipped systems run on it. A localization workflow moves every request through explicit nodes — parse, extract terminology, retrieve approved terms and translation memory, translate, optionally proofread, write back to the stores — so each step can be observed, skipped, or improved on its own. A retrieval platform compiles a StateGraph where language detection, query decomposition, multi-rewrite expansion, ensemble retrieval, reranking, and parent-document resolution are separate nodes with conditional edges — a query already in the corpus language skips translation, a simple one skips decomposition — and every model-calling node validates structured output and carries a fallback. Both have moved with the framework: the pipeline runs on the current LangChain/LangGraph v1 ecosystem, migrated as the interfaces changed rather than pinned to where they started.

Past the pipeline work sits the agent side of the framework: checkpointed state across conversation turns, middleware for summarization and task planning, and a human-in-the-loop interrupt-and-resume path — built and tested, ready for the first tool that should require a person's approval before it acts. That research led directly into the agent platform I am building now, where agents run as LangGraph graphs on a self-hosted Agent Protocol server, PostgreSQL holds threads and checkpoints, and any team's agent is a stored, strictly typed configuration rather than a separate codebase. LangSmith tracing runs over the production graphs, so the framework's structure pays off twice — once in design clarity, once in observability. The systems are listed in the Related Projects and Related Experience panels above.