LangGraph for Fraud Detection: Guardrails, Evals and Langfuse

Takeaways

• The old agent was a roughly 2,000-line prompt plus about 30 tools; results were non-deterministic and could take 4 minutes.
• LangGraph lets him run a directed acyclic graph that mixes deterministic JavaScript functions with LLM nodes, which he says cut runs to 10 to 20 seconds, exportable as a Mermaid chart.
• A human approves the agent's findings before anything changes on an account, and labeled simulated data feeds evals per workflow so he can see changes over time.
• Langfuse gives visibility into where time goes after his own tracing kept crashing; he calls LangSmith another good option that tends to be cloud-first.
• He argues LangGraph suits enterprises because of its guardrails, where a local agent that can do anything on disk carries too much risk.