AI Frameworks and Orchestration

This topic focuses on the framework implementation layer. The Agent and RAG topics discuss principles and tradeoffs, while Tools covers protocols separately. Here, the question is how different frameworks implement those principles and protocols—and what it costs to migrate when a framework no longer fits a project.

The material is organized into six modules. The LangChain ecosystem retains the existing LangChain/LangGraph paths. The remaining modules cover LlamaIndex, DSPy, Semantic Kernel, lightweight agent frameworks, and cross-framework selection and portable architecture.

"Lightweight" is a reading category, not a claim that AutoGen or CrewAI necessarily has a smaller runtime, fewer dependencies, or lower operating costs. Compare tool-use correctness, recovery semantics, latency, and cost on the same task rather than treating a framework's name as a capability guarantee.

Modules

1. LangChain Ecosystem (Chapters 1–13)—Chain/LCEL, building agents, state orchestration with LangGraph, and production feedback with LangSmith
2. LlamaIndex Ecosystem (Chapters 14–15)—data and index abstractions, query engines, and event-driven Workflows
3. Declarative Optimization with DSPy (Chapters 16–17)—declarative programming with Signature/Module, compilers, and optimizers
4. Enterprise Orchestration with Semantic Kernel (Chapters 18–19)—Kernel/Plugin/Planner, Process Framework, and Agent Framework
5. Lightweight Agent Frameworks (Chapters 20–21)—multi-agent abstractions in AutoGen and CrewAI, and PydanticAI's type-safe approach
6. Framework Selection and Portable Architecture (Chapters 22–23)—cross-framework technical comparison, identifying lock-in, and migration strategies

Where this topic fits

flowchart TB
    subgraph L3["Application architecture layer: principles"]
        AGENT["Agent topic"]
        RAG["RAG topic"]
        TOOLS["Tools topic"]
    end
    subgraph L4["Framework implementation layer: this topic"]
        LC["LangChain / LangGraph<br/>General-purpose agent orchestration"]
        LI["LlamaIndex<br/>Data and context"]
        DS["DSPy<br/>Declarative optimization"]
        SK["Semantic Kernel / Microsoft Agent Framework<br/>Enterprise orchestration and migration"]
        LW["AutoGen / CrewAI / PydanticAI<br/>Lightweight agent frameworks"]
        SEL["Selection and portable architecture<br/>A shared basis for comparing frameworks"]
    end
    AGENT --> LC
    AGENT --> LW
    RAG --> LI
    TOOLS --> LC
    TOOLS --> SK
    LC --> SEL
    LI --> SEL
    DS --> SEL
    SK --> SEL
    LW --> SEL

Technical connections between modules, not a product catalog

The six modules are not independent product introductions. Each revisits the same engineering dimensions: state models, persistence, tool contracts, evaluation and observability, and lock-in risk.

Question	Start here	How to extend the comparison
How is state updated and recovered?	LangGraph (Chapter 10), LlamaIndex Workflows (Chapter 15)	Chapters 19 and 22 compare state ownership, merging, and recovery semantics; similar diagrams do not imply equivalent runtimes
How do multiple agents collaborate?	SK (Chapter 19), AutoGen / CrewAI (Chapter 20)	Chapter 21 compares boundaries around messages, tasks, and typed calls
How do models call tools?	The Function Calling chapter in the Tools topic	This topic covers registration, execution, and error handling in frameworks; similar schemas do not imply the same protocol or retry semantics
How does evaluation drive improvement?	LangSmith (Chapter 13), DSPy (Chapter 17)	The former supplies evaluation evidence for development and production; the latter searches program parameters. Chapter 22 separates evaluation from observation
Which assets are worth preserving across frameworks?	Chapter 23	Revisit each framework's state, tool, and operational constraints to estimate migration costs

Suggested reading paths

For a first read, follow Chapters 1–23 in numerical order. The LangChain directory is grouped by topic: after Chapter 8, read Chapters 9–10, then return to Chapter 11 for version evolution. Directory grouping is not chapter order.

• Only interested in LangChain/LangGraph: go directly to the LangChain ecosystem.
• Selecting technology for RAG or a knowledge base: LlamaIndex ecosystem → LangChain: Ecosystem and Evolution → Framework Selection and Portable Architecture.
• Improving prompts systematically rather than tuning them by hand: Declarative Optimization with DSPy.
• Working in .NET or maintaining an existing SK system: Enterprise Orchestration with Semantic Kernel. Java teams can start with Chapter 8: LangChain4j.
• Building multi-agent collaboration or prioritizing type safety: Lightweight Agent Frameworks.
• Choosing a framework or planning a migration: start with Framework Selection and Portable Architecture, then consult individual framework chapters as needed.

Back to the documentation topic index.