LangChain

LangChain is an open-source framework for building LLM-powered applications and agents. It belongs to the Agent App / Orchestration Framework layer, not the software delivery workflow layer.

Primary reference: https://docs.langchain.com/oss/python/langchain/overview

Where LangChain fits

flowchart TB
    A[LLM / Model] --> B[LangChain]
    B --> C[Tools]
    B --> D[Prompt / instructions]
    B --> E[Retriever / data sources]
    B --> F[Agent logic]
    F --> G[AI app or agent service]

LangChain answers:

How do I build an AI app or agent that connects models, prompts, tools, data, and application logic?

It does not primarily answer:

How should an AI coding agent manage specs, plans, approvals, and delivery artifacts?

Core concepts

Concept	Role
Model	The LLM or chat model used by the app
Prompt/instructions	How the app guides model behavior
Tools	Functions the model/agent can call
Retriever	Component that fetches relevant external knowledge
Agent	Model-driven loop that decides whether to respond or use tools
Middleware	Custom control around agent behavior
Messages/state	Conversation and execution context

Simple agent flow

flowchart LR
    A[User request] --> B[Prompt / agent]
    B --> C[Model]
    C --> D{Need tool?}
    D -->|Yes| E[Tool call]
    E --> B
    D -->|No| F[Response]

RAG app flow

flowchart TB
    A[User question] --> B[Retriever]
    B --> C[Documents / Vector DB]
    C --> D[Context]
    D --> E[Prompt]
    E --> F[LLM]
    F --> G[Answer]

When to use LangChain

Use LangChain when:

• You are building an AI application or backend service.
• You need tools, model calls, prompts, retrievers, or structured outputs.
• You want a fast way to build agents that call functions.
• You need integrations with model providers, vector stores, tools, and observability.
• Your workflow is application runtime behavior, not software delivery governance.

When not to use LangChain

Do not use LangChain as a replacement for:

Need	Better layer
Spec-first software delivery	Spec Kit or OpenSpec
Enterprise approval/audit lifecycle	AWS AI-DLC
Agent CLI/harness for coding in a repo	Codex CLI, Claude Code, Hermes
TDD/review discipline	Superpowers
Multi-phase coding project memory	GSD

LangChain and workflow frameworks

LangChain can be the app framework inside a repo whose development is governed by another workflow.

Example:

flowchart TB
    A[OpenSpec change] --> B[Implement LangChain app]
    B --> C[Retriever + tools + prompts]
    C --> D[Tests and evals]
    D --> E[Sync/archive specs]

In that stack:

• OpenSpec owns change artifacts.
• LangChain implements the AI app behavior.
• Superpowers can enforce TDD/review.
• CI/evals prove correctness.

Step-by-step: build a RAG chatbot

1. Define the feature with OpenSpec or Spec Kit.
2. Identify knowledge sources.
3. Create retriever/vector store.
4. Write prompt/instructions.
5. Create LangChain chain or agent.
6. Add citations or source references.
7. Add tests/evals for answer quality.
8. Add latency and cost monitoring.
9. Review and ship.

Definition of done

An AI app built with LangChain is done when:

1. Inputs/outputs are defined.
2. Tool permissions are explicit.
3. Retrieval behavior is tested.
4. Prompt behavior is evaluated.
5. Errors and fallbacks are handled.
6. Cost and latency are measured.
7. Observability exists for production use.