​
The Rise of Agent Frameworks

Building a sophisticated AI agent from scratch is a significant engineering effort. It requires managing communication with the LLM, maintaining conversational memory, integrating external tools, and orchestrating complex, multi-step logic.4 Agent frameworks exist to simplify this process by providing structured, reusable components for these common tasks.2 They allow developers to build applications that can reason, plan, and act autonomously to achieve goals. However, the power and flexibility of these frameworks can also be a source of complexity. Frameworks like LangChain, with their vast array of modules and chains, can have a steep learning curve and lead to code that is difficult to debug and maintain, a problem some developers have described as “painful” when trying to scale.18 This very complexity creates a need for a higher-level architectural guide to inform how these powerful tools should be used.

​
LangChain & LlamaIndex: A Tale of Two Philosophies

Among the many frameworks, LangChain and LlamaIndex are arguably the most prominent, and their differing philosophies perfectly illustrate the need for an overarching architecture like CWA.
  • LangChain: LangChain is a highly flexible, general-purpose orchestration framework. It is often described as a “Swiss Army knife” for building LLM applications.20 Its core strength lies in its modularity, allowing developers to “chain” together LLM calls with tools, memory systems, and other components to create complex, agentic workflows.3 It provides a vast library of integrations for different models, databases, and APIs, giving developers maximum control.3 LangChain’s components can be used to implement nearly every layer of the CWA.
  • LlamaIndex: LlamaIndex, by contrast, is a more specialized, data-centric framework. It excels at the “data-to-context” part of the problem.22 Its primary focus is on providing a highly optimized and streamlined experience for indexing data from diverse sources and performing sophisticated retrieval for RAG applications.18 While LangChain is about general-purpose orchestration, LlamaIndex is a best-in-class implementation toolkit for CWA Layer 3.22
A very common and powerful pattern is to use LlamaIndex and LangChain together.21 In this pattern, LlamaIndex is used for its superior data indexing and retrieval capabilities (to populate Layer 3), and LangChain is used to orchestrate the overall agentic logic, tool use, and memory (to manage the other CWA layers). The fact that developers are naturally combining these tools is powerful evidence that they might actually be trying to build an architecture that separates data-handling concerns from agent-orchestration concerns—a separation that CWA makes explicit and formal.

​
CWA as the Framework-Agnostic Architectural Guide

This leads to the central argument of this subsection: CWA is the framework-agnostic architectural blueprint that guides the effective use of toolkits like LangChain and LlamaIndex. Instead of a developer starting a project by asking “Which LangChain module should I use?”, CWA encourages them to start by asking “What architectural layers does my application need?” The developer first thinks architecturally: “To solve this business problem, my AI needs a persistent identity (Layer 1), it needs to know about the user (Layer 2), it must be grounded in our product documentation (Layer 3), and it needs to guide the user through a three-step process (Layer 4).” Only after defining these architectural requirements does the developer turn to the framework. The question then becomes much more focused: “What is the best LangChain component to implement Layer 2? What is the best way to use LlamaIndex to implement Layer 3?” CWA provides the structured thinking that prevents developers from getting lost in the complexity of the frameworks. It turns framework selection and usage from a bottom-up, tool-driven process into a top-down, architecturally-driven one. This leads to cleaner, more modular, and more maintainable code because the code’s structure directly reflects the application’s architectural design.

​
Mapping CWA Layers to Agent Framework Concepts

To make this relationship concrete, the following table helps by translating the conceptual CWA layers into specific, implementable components from popular agent frameworks. This provides a practical guide for developers looking to apply the CWA pattern in their work.
CWA LayerLangChain Component(s)LlamaIndex Component(s)Other Frameworks (Examples)
1. InstructionsSystemMessage in Prompt Templates, Agent prompt customization. 3system_prompt in Query Engines.Botpress: Instructions in the visual builder. 4 AutoGen: Role definitions in ConversableAgent. 17
2. User InfoCustom logic feeding user data into prompt templates. Can be managed within ConversationBufferMemory. 3Metadata filtering on indexes (e.g., user_id).No-Code Tools (Lindy): Context window setup with CRM integrations. 24
3. Curated Knowledge ContextDocumentLoaders, VectorStoreRetriever, create_retrieval_chain. 3DataReader (via LlamaHub), VectorStoreIndex, RetrieverQueryEngine. This is LlamaIndex’s core strength. 3Haystack: Pipelines with Retriever and Reader nodes. 13
4. Task/Goal State ContextAgents (e.g., ReAct, Self-Ask), Chains (e.g., SequentialChain), LangGraph for stateful graphs. 3QueryPlanTool for decomposition, SubQuestionQueryEngine.AutoGen: GroupChatManager orchestrating multiple agents. 17 CrewAI: Defining Tasks and Processes. 24
5/6. Chat History & SummaryMemory modules (e.g., ConversationBufferWindowMemory, ConversationSummaryMemory). 3ChatMemoryBuffer, CondenseQuestionChatEngine.All Frameworks: Most frameworks provide some form of memory management as a core feature. 2
7/8. Tools & Function ResultsTools / Toolkits, AgentExecutor for the ReAct loop, create_tool_calling_agent. 2FunctionTool, QueryEngineTool.Haystack: Integration with external APIs via custom nodes. 5 AutoGen: register_function. 5
9. Few-Shot ExamplesFewShotPromptTemplate, providing examples directly in the prompt. 6Examples provided in the prompt template for a query engine.Prompt Engineering: This is a general technique applicable across all frameworks. 6
10. Dynamic Output FormattingOutputParser classes (e.g., PydanticOutputParser, JsonOutputParser), with_structured_output function. 1OutputParser modules, response synthesis customization.All Frameworks: Most frameworks that support function/tool calling also support structured output specifications like JSON Schema.