# Interrupt NYC: Opening Keynote

## Executive summary

Harrison Chase outlines the shift in AI development from simple LLM prototypes to complex, domain-specific agent applications. The core concept is 'owning your intelligence' by building specialized components around foundational models. LangChain addresses this through its platform, LangSmith, which provides three pillars: an open, model-neutral harness (like LangGraph or Deep Agents); a compounding loop for continuous learning from user interactions; and robust governance for internal enterprise use. Key product announcements include Managed Deep Agents, the decision model Jev, and LangSmith Engine v2, which automates issue detection and red teaming.

## Key takeaways

- The Shift to Domain-Specific Agents: AI applications are moving beyond simple LLM calls. Success requires building specialized harnesses and data layers around the model to provide differentiated, domain-specific value (e.g., Rogo for finance, Harvey for legal).
- Three Pillars of Owning Intelligence: To own intelligence, organizations must focus on: 1) An open, model-neutral harness (e.g., LangGraph, Deep Agents); 2) Compounding intelligence from user usage (the feedback loop); and 3) Governance (managing auth, costs, and internal access).
- LangSmith Engine v2 Capabilities: The platform's intelligence layer, LangSmith Engine v2, can now proactively test its fixes on deployments, perform red teaming simulations, and automate the creation of initial evaluation datasets, having scanned over 70 million traces and detected over 21,000 issues.

## Technical details

- Managed Deep Agents: This platform combines the agent harness and managed infrastructure into one, simplifying deployment. New features include enhanced auth primitives (user-level vs. service-level identity), user-level memory, and built-in web search via Parallel.
- Decision Models (Jev/SemIf): Decision models, such as Jev, are specialized tools that make decisions (scoring questions, choosing categories) rather than generating text. They are critical for model routing, guardrails, and evaluating agent performance. They are accessible via the LangSmith LLM Gateway.
- LangSmith LLM Gateway: This gateway normalizes LLM access (e.g., OpenAI and Anthropic message format) and is crucial for governance. It enables cost control, rate limiting, and implementing stateful fallbacks across multiple model providers.
- Observability and Evals: LangSmith Trajectories standardize the representation of multi-turn agent interactions, making debugging faster and providing a standard format for fine-tuning data. LangSmith Fine-Tuning uses the smithtune CLI to select, pre-process, and pass data to training infrastructure (e.g., Fireworks, Baseten).
- Agent Runtime Components: An agent requires three components: 1) Business Logic (instructions, tools); 2) The Harness (e.g., Deep Agents, LangGraph) to manage context; and 3) Infrastructure (durable execution, secure tool access, model access). LangSmith Deployment addresses the infrastructure layer.

## Practical implications

- Build engineers can accelerate time-to-value by using Managed Deep Agents, which pre-assemble the complex agent harness and infrastructure.
- Implementing a centralized LLM Gateway is crucial for enterprise governance, allowing for cost control, rate limiting, and failover across multiple model providers.
- The adoption of LangSmith Trajectories simplifies the MLOps process by standardizing agent interaction data, making it easier to fine-tune models and debug complex agent flows.
- Engine v2 allows teams to shift from reactive debugging to proactive testing and red teaming, significantly improving agent reliability before production deployment.

## Topics

AI Agents, LLM Orchestration, MLOps, Build Engineering, Governance, LangChain, LangSmith, LangGraph, Deep Agents, Jev, LangSmith LLM Gateway, LangSmith Trajectories, smithtune CLI, LangSmith Engine v2, SmithDB

Source: https://www.youtube.com/watch?v=950byF7njfw