Watch our recorded sessions from Interrupt here: https://interrupt.langchain.com/video/?utm_medium=social&utm_source=youtube&utm_campaign=q2-2025_interrupt-2025_co
LangChain CEO Harrison Chase reveals why quality remains the #1 blocker for getting AI agents into production—and introduces a systematic 3-stage evaluation framework to solve it.
Based on a survey of agent builders, quality outranks cost and latency as the biggest barrier to production deployment. While prototypes may work for demos, production systems need far higher reliability. Harrison introduces eval-driven development as the solution to bridge this critical gap.
What You'll Learn:
- The three types of evaluations: Offline, Online, and In-the-Loop evals explained
- How LangSmith transforms production traces into custom evaluation datasets
- When to use LLM-as-judge vs. deterministic evaluators for your specific use case
- New launches: Chat simulations, eval calibration, and open-source OpenEvals package
Harrison demonstrates how "great evals start with great observability" and why evaluation must be treated as a continuous journey rather than a one-time task. From offline testing with curated datasets to real-time production monitoring, learn the complete evaluation lifecycle that successful agent builders implement.
Featured Products:
LangSmith's unified observability and evaluation platform, plus the new open-source OpenEvals package with pre-built evaluators for code, RAG, extraction, and tool calling.
#LangChain #AIAgents #LangSmith #Evaluation #ProductionAI #AgentDevelopment