Architectural Shift

• LinkedIn moved from hard-coded branching to a centralized architecture using LangGraph, selecting it after evaluating 89 frameworks because it integrated with their existing LangChain infrastructure without requiring a rewrite.5:40
• The system now uses a plan-execute-replan loop, which the team asserts is significantly more robust than previous methods for handling the iterative, stateful nature of hiring.4:01

Design Constraints

• The engineering team rejected LangGraph’s native interrupt-style pause/resume flow because recruiter behavior—which often involves switching topics or stepping away—proved too unpredictable for rigid checkpoints.13:17
• They shifted to a context-driven human-in-the-loop model where every user input triggers an end-to-end graph execution, while only the minimum necessary cross-turn context is persisted to keep system state manageable.13:42

Evaluation and Reliability

• Because LinkedIn company policy currently restricts sending full production data to LangSmith, the team built a custom internal mirroring system to trace interactions and evaluate performance using human annotation alongside LLM-as-judge techniques.11:02
• To manage the probabilistic nature of LLMs, the team focused heavily on harness engineering—including template confirmation, one-shot tool guards, and programmatic response assembly—to ensure consistency in recruiter-facing actions.15:02