Limits of Linear LLM Pipelines

The first generation of LLM applications were built as chains.

The Chain Architecture

A chain is a sequence of deterministic steps where the output of one step becomes the input of the next.

graph LR
    Input --> PromptTemplate
    PromptTemplate --> LLM
    LLM --> OutputParser
    OutputParser --> FinalOutput

Why Chains Work for Simple Tasks

For tasks like "Summarize this article" or "Translate this text", chains are perfect.

1. Predictable: Step A always leads to Step B.
2. Fast: No complex routing logic.
3. Easy to Debug: If it fails, you know exactly which step broke.

Where Chains Fail

Real-world problems are rarely linear. Consider a "Travel Planning Assistant".

• User: "Book me a flight to Paris."
• Chain: Calls Flight API.
• API: Returns "No flights found".
• Chain: Crashes or returns the raw error to the user.

A linear chain has no mechanism to say: "Oh, no flights? Let me try checking nearby airports or different dates."

The Need for Cycles

To handle failure or complexity, you need cycles (loops). You need the ability to go back to a previous step and try again with new information. Linear DAGs (Directed Acyclic Graphs) cannot loop. This is the fundamental limit that LangGraph overcomes.