AI [Multi-]Agent Frameworks
LangChain - building LLM-based context-aware apps from RAG to core toolkit for building individual agents in more advanced frameworks like CrewAI and LangGraph. Uses: quickly create LLM-based apps w/smart agent behavior, connect LLMs to outside info and tools to form simple AI agents.
LangGraph - extension of LangChain to build agentic systems using a graph structure - cyclical and stateful workflows, where agents loop, self-correct, make decisions based on current state of the process - more complex and robust interactions than chains. Uses: complex smart agent systems w/control over moving between steps and repeating actions, chat AI remembering long conversations that can follow different paths, systems where agents’ subsequent actions depend on previous ones.
Google ADK — Google’s agent framework, A2A-native.
Claude Agent SDK — the harness behind Claude Code (tools, permissions, sub-agents).
OpenAI Agents SDK — successor of Swarm: agents, handoffs, guardrails, sessions, built-in tracing.
smolagents (HF) — minimalist code-acting agents (agent writes Python instead of JSON tool calls).
LlamaIndex - connecting LLMs to custom data sources. Core strength - build RAG apps, agents are often specialized for complex data querying and synthesis. Uses: build generative AI apps by connecting LLMs with different types of data (like docs or DBs), develop systems that find info and then generate text, manage data for complex AI solutions that need smart ways to take in and ask questions about data.
AutoGen - building apps w/LLM agents who talk to each other to solve tasks: automating complex workflows like code generation, execution, and human feedback. Uses: automate difficult software tasks (write code, find errors, test code), chat AI w/many smart agents, develop AI agents that use tools and run code on the fly.
Microsoft Agent Framework created by converging Semantic Kernel + AutoGen.
CrewAI - orchestrating role-playing, autonomous AI agents into teams to accomplish a shared goal. Uses: organize AI agents into teams for specific tasks (marketing w/researcher, writer, and editor), automate business steps by assigning roles, building specialized AI systems acting like human teams.
Mesa (Python) - Agent-based modeling and simulation of complex systems to understand the emergent behavior of many simple agents. Uses: simulate people crowd behavior (fake news spreading), animal groups or economies, how agents interact over time.
Ray (Python) - scaling AI and Python apps - distributing workload of many agents across a cluster, massive parallelism for training or real-time inference. Uses: training complex AI models, controlling groups of self-driving cars or drones, building scalable ML services.
Other - PydanticAI — type-safe agents w/validated structured outputs. DSPy — declarative LLM programs w/automatic prompt optimization.
Source: https://cloud.google.com/discover/what-is-a-multi-agent-system
LangChain vs. LangGraph
LangChain provides a stateless abstraction layer for chaining LLM operations in a sequential DAG workflow, designed for deterministic, step-by-step pipelines such as retrieve → summarize → answer. Core components:
DocumentLoaders (data retrieval), TextSplitters, Prompts, LLM wrappers, Chains, Agents, and Memory.
Each stage executes once, passing outputs downstream.
Limited long-term state — memory components can preserve short context but not a persistent global state.
LangChain excels at linear, composable LLM pipelines where the task order is predefined.
LangGraph, in contrast, is a stateful graph-based orchestration layer within the LangChain ecosystem, specialized for multi-agent, nonlinear, and iterative workflows.
Nodes represent actions (e.g., add task, complete task, summarize), edges define transitions, and a shared State object maintains global context accessible to all nodes.
Allows loops and revisiting prior states, enabling context-aware, adaptive, and conversational agents.
Especially suited for task-oriented assistants or agent systems that must handle evolving inputs, dynamic routing, and persistent memory.