FULL STORY

Jev-as-a-Judge: From Blog Post to LangSmith

LangChain introduced Jev-as-a-Judge for agent evals and quickly shipped it as a LangSmith integration, enabling low-cost evaluation of individual production traces.

2026-09-20 ~ 2026-09-22 · 2 episodes · 11 posts

Episode 1 · LangChain Launches Jev-as-a-Judge for Agent Evals, Now Live in LangSmith (2026-09-20, 9 posts)

On September 20, the LangChain team released the guide "Jev-as-a-Judge for Agent Evals" (written by Daniel Shea and Seán Roche, with co-founder Harrison Chase and swyx sharing it), proposing a brand-new evaluator called Jev as a replacement for LLM-as-a-Judge in online agent evaluation.

Confirmed

  • Jev is an evaluator that differs fundamentally in mechanism: it directly returns structured, typed answers instead of first generating text and then parsing it like an LLM judge does, thereby avoiding the uncertainty introduced by text parsing
  • LangChain's benchmark data: scoring variance is 92-913x lower than LLM judges, speed is up to 5x faster, cost savings reach up to 99%, with each evaluation costing about $0.34
  • Team member Sydney Runkle published a separate blog post, "Building a Harness with Jev," covering hands-on experience using Jev to build evaluation tooling within agent loops (LLM decision → tool execution → model evaluation), and added further usage details not covered in the blog via follow-up shares

Why it matters

  • LLM-as-a-Judge has long been plagued by unstable scoring, high costs, and slow speeds; Jev's structured-output approach offers a cheaper, faster, and more stable alternative for semantic verification, making it especially well suited for continuously evaluating agent performance in production

Episode 2 · LangSmith Launches Jev-as-a-judge for Cheap Full-Trace Evaluation (2026-09-22, 2 posts)

LangChain launched Jev-as-a-judge on LangSmith, enabling full evaluation of every production trace at low cost, with more judging criteria per trace and evals generated directly from large-scale trace mining.