Architecture

Layered Design

1. Primitives

   • runnables
   • prompts
   • output parsers

2. Model Adapters

   • provider-specific wrappers implementing Model
   • ModelRunnable bridge for chain usage

3. Agent Runtime

   • instruction interpretation
   • tool execution
   • memory context propagation

4. Applications

   • playground and downstream products

Execution Paths

Chain Path

input -> runnable composition -> parsed structured output

Agent Path

user input -> model nodes -> tool calls/final response -> loop until final

Why This Structure

• Keeps primitives reusable outside agent runtime.
• Allows deterministic testing of chain stages.
• Keeps provider code isolated from orchestration logic.