LLM Agent Architecture for Beginners
LangChain
CrewAI
Dify
LlamaIndex
Meta
LM Studio
The article, based on a talk at an internal meetup, explains how a raw LLM server evolves into a full agent system. It covers key components: chat context storage, internet search, RAG for local data, tools, system prompts, orchestration, and the MCP protocol.
The article uses the metaphor of a fish named Dory to describe a bare LLM server, which lacks short-term memory and the ability to act. To make it useful, several layers are added: chat context storage for conversation history; internet search to ground answers in current data; local retrieval via RAG to incorporate private knowledge; tools that allow the model to request actions, executed by the outer system; system prompts for consistent behavior; and orchestration with a plan-do-check cycle for autonomy. The MCP protocol standardizes tool interfaces, and the author notes that while models are becoming more capable, they remain simple at their core. The article concludes by listing frameworks and tools like LangChain, CrewAI, Dify, LlamaIndex, and others that build on these principles.
- Abbreviations
- LLM = Large Language Model — Большая языковая модель
- RAG = Retrieval-Augmented Generation — Генерация с дополнением извлечением
- MCP = Model Context Protocol — Протокол контекста модели
- IDE = Integrated Development Environment — Интегрированная среда разработки
- CLI = Command-Line Interface — Интерфейс командной строки
- PDCA = Plan-Do-Check-Act — Планируй-Делай-Проверяй-Действуй
Source: Habr — хаб ML —
original