Reflection Loops in LangGraph: Write, Critique, Revise

by bam

You can build a LangGraph graph and call a model, but you haven't yet built an agent that judges its own output, revises it, and knows when to stop. This course builds that write, critique, revise loop one piece at a time on a running support-reply example, then hardens it: bounded termination so it always exits, a data rubric an LLM can actually score, LangSmith tracing that makes bugs legible, and an eval that reuses the very same rubric. Module 5 treats reflection as a budget on a cost and quality curve, and the capstone redeploys the whole stack on a brand-new domain (rewriting dense legal clauses into plain language) to prove the pattern transfers. It's TypeScript first with a Python translation track, grounded in the open-source wanderlearn-field-reporter repo, and every example runs offline with no API key. Follow along in the wanderlearn-field-reporter repo, and see the loop running in the deployed app.