Candidly and LangSmith jointly released a state perception agent and self-evolving engine

On August 26, 2026, the LangChain Blog announced two advancements: First, Candidly’s agent Cait used LangSmith for annotation pipelines to infer user states and guide responses by analyzing dialogue tracking records, achieving 92.3% human-like consistency in this process. Second, LangSmith introduced a technical engine called LangSmith Engine, designed to analyze large-scale tracking data, transform repetitive failures into actionable issues, identify operating patterns, and generate evaluations and repair solutions, thereby enhancing the agent’s own optimization capabilities.