An Evolving DAG for the LLM world - Julia Schottenstein of LangChain at Small Data SF 2024

Building Agentic Systems with LangChain: From DAGs to Directed Cyclic Graphs

LangChain has emerged as a popular open-source framework for Python and TypeScript developers looking to build agentic systems that combine the power of Large Language Models (LLMs) with organizational data. The framework addresses a fundamental challenge in AI application development: while LLMs possess incredible capabilities, they lack context about specific businesses, applications, and recent events beyond their training cutoff dates.

Augmenting LLMs with Context and Tools

The core value proposition of LangChain lies in helping developers augment LLMs with:

• Private documents and data for domain-specific reasoning
• Tool usage through APIs with defined instructions
• Up-to-date information to overcome training data limitations

This augmentation typically happens through chains - discrete, ordered steps similar to directed acyclic graphs (DAGs) in data pipelines. The most common implementation is the Retrieval-Augmented Generation (RAG) chain, where:

1. A question enters the system
2. Relevant context is retrieved from a vector database
3. The question, context, and prompt instructions are sent to the LLM
4. The LLM generates a contextually-aware response

The Evolution from Chains to Agents

While chains provide reliable, predetermined workflows, the future of AI applications increasingly points toward agents. In technical terms, an agent represents a system where the LLM decides the control flow dynamically rather than following predefined code paths. This fundamental shift transforms traditional DAGs into directed graphs that can include cycles.

The ability to iterate and learn from failures becomes crucial in agent design. Unlike deterministic code that produces identical results on repeated execution, LLMs can improve their performance on subsequent attempts by understanding what went wrong in previous iterations. This capability is exemplified by sophisticated code generation agents that:

• Reflect on problems before execution
• Generate and test multiple solution approaches
• Iteratively refine outputs based on test results
• Create cycles in their execution graphs for continuous improvement

Key Challenges in Building Reliable Agents

Balancing Flexibility and Control with LangGraph

LangGraph represents LangChain's solution to the flexibility-reliability trade-off. This orchestration framework introduces several key innovations for agent development:

Real-World Agent Applications

Several production agents demonstrate the practical applications of these concepts:

• Roblox Studio AI creates entire virtual worlds from natural language prompts, generating scripts and assets automatically
• TripAdvisor's travel agent builds personalized itineraries based on user preferences, group size, and travel dates
• Replit's coding agent generates code, creates tests, and automates pull request creation

These applications showcase how agents can move beyond simple chat interfaces to become sophisticated task-completion systems that understand context, iterate on solutions, and deliver tangible value to users.

The Future of AI Orchestration

The evolution from traditional DAGs to directed cyclic graphs represents a fundamental shift in how we approach AI application development. While DAGs remain valuable for deterministic data pipelines, the ability to incorporate cycles opens new possibilities for building intelligent systems that can plan, reflect, and improve through iteration. As agent technology continues to mature, frameworks like LangChain and LangGraph provide the necessary tools to build reliable, flexible, and powerful agentic applications that can tackle increasingly complex real-world problems.