LangGraph and LangChain: LinkedIn hiring agent architecture

LinkedIn's hiring team explains how it built an AI agent that the speakers say cut time to interview by 60% for small businesses. The workflow collects hiring requirements, drafts a job description, sources candidates and evaluates applicants against recruiter-defined qualifications.

The architecture moved from hard-coded branches to sequential LangChain chains, then to a LangGraph central planner with a plan-execute-replan loop. The speakers say LinkedIn selected LangGraph after evaluating 89 frameworks because it fit existing infrastructure and reused LangChain primitives without a rewrite.

A practical lesson concerns human approval. Recruiters may ignore a suggestion or change topics, so the team replaced an interrupt-based approach with a graph run for each input. It retains minimal context between turns, allowing the planner to confirm, reject or move past a pending suggestion.

The talk also covers conversational memory, state checkpoints and controls for consistent output and tool execution. LangSmith supports tracing and evaluation, though LinkedIn policy limits which production data the team can send to it. The team describes human annotation, LLM judging and manual prompt refinement. This is an architecture overview, with no local deployment or hardware setup instructions.