ARTICLE

Agent Frameworks

Updated 2 May 2025

agentsframeworkslangchaincrewailanggraphorchestration

Agent Frameworks

You understand what an agent is — an LLM in a loop, reasoning and acting until a task is done. But you don’t build agents from scratch. You use a framework — an agent harness — that handles the plumbing: tool routing, memory management, error handling, orchestration.

Think of it like web frameworks. You could build a web server from raw TCP sockets. But you use Express, or SvelteKit, or Django, because the patterns are solved. Agent frameworks do the same for AI.

The landscape is moving fast. Here’s what matters right now.

The Major Frameworks

LangChain / LangGraph

The ecosystem. LangChain is the most widely used AI application framework. LangGraph (its agent-specific layer) gives you fine-grained control over agent workflows.

Aspect	Details
Best for	Custom agent workflows, complex multi-step pipelines
Architecture	Graph-based. You define nodes (actions) and edges (transitions)
Language	Python, TypeScript
Strengths	Flexibility, ecosystem size, LangSmith observability
Limitations	Steep learning curve, over-abstracted for simple cases
Who uses it	Most AI application developers start here

LangGraph is where the interesting work happens. It lets you build stateful, multi-actor workflows where you control exactly when the LLM reasons, when tools execute, and when humans approve.

CrewAI

Multi-agent orchestration. CrewAI’s model is intuitive: define agents with roles, give them tools, assign tasks, and let them collaborate.

Aspect	Details
Best for	Multi-agent collaboration, task delegation
Architecture	Role-based. Agents have backstories, goals, and tools
Language	Python
Strengths	Intuitive API, quick to prototype, built-in delegation
Limitations	Less control than LangGraph, newer ecosystem
Think of it as	“Hire a crew of specialists and let them figure it out”

AutoGen (Microsoft)

Multi-agent conversation. AutoGen treats agent collaboration as a conversation between specialised agents who debate, critique, and build on each other’s work.

Aspect	Details
Best for	Research, complex problem-solving, code generation with review
Architecture	Conversational. Agents talk to each other
Language	Python
Strengths	Good for tasks that benefit from debate/review cycles
Limitations	Can be verbose, hard to control cost

Native Tool Use APIs

Sometimes you don’t need a framework at all. The model providers give you tool use directly:

Provider	What They Offer
Anthropic	Tool use API + MCP (Model Context Protocol). Clean, well-documented.
OpenAI	Assistants API. Managed threads, file search, code interpreter.
Google DeepMind	Gemini function calling. Tightly integrated with Google services.

When to use native APIs: Simple tool-use patterns, single-agent workflows, or when you want minimum abstraction.

When to use a framework: Complex orchestration, multi-agent, stateful workflows, or when you need observability.

Agent Design Patterns

Frameworks implement common patterns. Understanding the patterns helps you choose the right tool:

ReAct (Reason + Act)

The simplest useful pattern. The agent thinks, acts, observes, repeats.

Thought: I need to find the company's revenue
Action: search("NVIDIA 2024 revenue")
Observation: $60.9 billion
Thought: Now I can answer the question
Action: respond("NVIDIA's 2024 revenue was $60.9B")

Use when: Single-agent, straightforward tasks. Most coding assistants use this.

Plan and Execute

The agent creates a plan first, then executes each step. Better for complex tasks where the sequence matters.

Plan:
1. Search for the company's financials
2. Compare to competitors
3. Summarise trends
4. Write report

Execute: [runs each step, adjusting the plan if needed]

Use when: Multi-step research, report generation, complex analysis.

Multi-Agent

Specialised agents collaborate. A planner agent decomposes the work. Worker agents handle subtasks. A reviewer agent checks quality.

Use when: Tasks that benefit from different “perspectives” or expertise. Code generation + review is a natural fit.

Human-in-the-Loop

The agent works autonomously but pauses at critical points for human approval. The responsible default for anything high-stakes.

Use when: Deploying AI where errors have real consequences. The alignment-aware choice.

Model Context Protocol (MCP)

MCP deserves special mention. Created by Anthropic, MCP is becoming the universal standard for connecting AI to tools and data.

Instead of building custom tool integrations for every framework and model, you build one MCP server. Any MCP-compatible client (Claude, Cursor, many IDE tools) can use it.

Think USB, not proprietary cables. One standard. Infinite tools.

MCP is why the agent ecosystem is accelerating — the tool integration problem is being solved at the protocol level.

How to Choose

Your situation	Start here
First agent project	Native tool use API (Anthropic or OpenAI)
Need custom workflows	LangGraph
Want multi-agent fast	CrewAI
Building a product	LangGraph + LangSmith (for observability)
Just experimenting	Ollama + any framework locally

The honest truth: most agent tasks don’t need a framework at all. A well-prompted model with tool access handles 80% of use cases. Frameworks shine when you need orchestration, state management, or multi-agent coordination.

What I’m Still Learning

How to evaluate agent performance systematically (it’s not like benchmarking a model)
The right level of abstraction — too much framework and you can’t debug, too little and you’re reinventing wheels
How MCP changes the framework landscape as it matures

Go Deeper

AI Agents — The concepts and theory
RAG & Retrieval — Agents often use RAG for knowledge grounding
AI Software & Tools — The broader developer ecosystem
AI Models — The models that power these agents
Prompt Engineering — How you instruct agent behaviour
AI Security — Security considerations for autonomous agents
AI Intelligence Hub — Back to the hub home

Sources

LangChain Docs — Framework documentation
LangGraph — Agent workflow engine
CrewAI — Multi-agent framework
Anthropic MCP — Model Context Protocol
Anthropic Tool Use — Native tool use docs
Anthropic “Building Effective Agents” — Best practices guide

AI Agent frameworks

	smolagents	The simplest way to build great agents. Agents write python code to call tools and orchestrate other agents. Priority support for open models like DeepSeek-R1!
	YoMo	Stateful Serverless LLM Function Calling Framework with Strongly-typed Language Support
	DeepSeek MCP Server	Model Context Protocol server for DeepSeek’s advanced language models.
	SuperAgentX	SuperAgentX: A Lightweight Open Source AI Framework Built for Autonomous Multi-Agent Applications with Artificial General Intelligence (AGI) Capabilities.
	Anda	A Rust framework for AI agent development, designed to build a highly composable, autonomous, and perpetually memorizing network of AI agents.
	Daydreams	Daydreams is a generative crosschain agent framework for executing anything onchain. Autonomous and easy to build on enabling the next generation of agents.
	RIG	Build modular and scalable LLM Applications in Rust.
	Just-Agents	A lightweight, straightforward library for LLM agents - no over-engineering, just simplicity!
	Alice	An autonomous AI agent on ICP, leveraging LLMs like DeepSeek for on-chain decision-making. Alice combines real-time data analysis with a playful personality to manage tokens, mine BOB, and govern ecosystems.
	Upsonic	Upsonic offers a cutting-edge enterprise-ready agent framework where you can orchestrate LLM calls, agents, and computer use to complete tasks cost-effectively.
	AIMatrices	AIMatrices is a lightweight, high-performance, scalable, and open source AI application rapid building platform designed to provide developers with an efficient and convenient AI application development experience. It integrates multiple advanced technologies and tools to help users quickly build, deploy, and maintain AI applications without having to write complex code from scratch.

RAG frameworks

	RAGFlow	An open-source RAG (Retrieval-Augmented Generation) engine based on deep document understanding. It offers a streamlined RAG workflow for businesses of any scale, combining LLM (Large Language Models) to provide truthful question-answering capabilities, backed by well-founded citations from various complex formatted data.
	Autoflow	AutoFlow is an open-source knowledge base tool based on GraphRAG (Graph-based Retrieval-Augmented Generation), built on TiDB Vector, LlamaIndex, and DSPy. It provides a Perplexity-like search interface and allows easy integration of AutoFlow’s conversational search window into your website by embedding a simple JavaScript snippet.
	DeepSearcher	DeepSearcher combines powerful LLMs (DeepSeek, OpenAI, etc.) and Vector Databases (Milvus, etc.) to perform search, evaluation, and reasoning based on private data, providing highly accurate answer and comprehensive report.
	KAG	KAG is a logical reasoning and Q&A framework based on the OpenSPG engine and large language models, which is used to build logical reasoning and Q&A solutions for vertical domain knowledge bases. KAG can effectively overcome the ambiguity of traditional RAG vector similarity calculation and the noise problem of GraphRAG introduced by OpenIE. KAG supports logical reasoning and multi-hop fact Q&A, etc.
	Youtu-GraphRAG	Youtu-GraphRAG proposes a new graph retrieval-augmented generation paradigm which vertically unifies graph construction, indexing and retrieval through schema. It advances GraphRAG towards enterprise-level applications, allowing seamless cross-domain adaptation with minimum human intervention, with (i) a dynamic schema-bounded extraction; (ii) a comprehensive community detection algorithm and (iii) a schema-guided query decomposition for iterative reflection and reasoning.

^ Back to Contents ^

RAG frameworks

	ATTPs	A foundational protocol framework for trusted communication between agents. Any agents based on DeepSeek, By integrating with the ATTPs SDK, can access features such as agent registration, sending verifiable data, and retrieving verifiable data. So that it can make trusted communication with agents from other platforms.
	translate.js	AI i18n for front-end developers. It can achieve fully automatic HTML translation with just two lines of JavaScript. You can switch among dozens of languages with a single click. There is no need to modify the page, no language configuration files are required, and it supports dozens of fine-tuning extension instructions. It is SEO-friendly. Moreover, it opens up a standard text translation API interface.
	agentUniverse	agentUniverse is a multi-agent collaboration framework designed for complex business scenarios. It offers rapid and user-friendly development capabilities for LLM agent applications, with a focus on mechanisms such as agent collaborative scheduling, autonomous decision-making, and dynamic feedback. The framework originates from Ant Group’s real-world business practices in the financial industry. In June 2024, agentUniverse achieved full integration support for the DeepSeek series of models.
	BotSharp	BotSharp is an open-source multi-agent application development framework. From simple chatbots to multi-agent collaboration and complex tasks like the Text To SQL framework, it provides out-of-the-box solutions to quickly integrate large model capabilities into existing business systems. It also includes built-in knowledge base and session management features. The framework has been thoroughly tested with DeepSeek V3 models, and thanks to the performance of DeepSeek V3, the framework’s performance is on par with other proprietary models.
	Eino	Eino (pronounced like “I know”) aims to be the best LLM application development framework in the Go language. It draws on the design concept of excellent LLM frameworks in open source communities such as LangChain and LlamaIndex, while absorbing cutting-edge research results and practical application experience, providing a LLM application development framework that is more in line with Go programming conventions, emphasizing simplicity, scalability, reliability and efficiency.
	Youtu-Agent	Youtu-Agent is a flexible, high-performance framework for building, running, and evaluating autonomous agents. Beyond topping the benchmarks, this framework delivers powerful agent capabilities, e.g. data analysis, file processing, and deep research, all with open-source models.

^ Back to Contents ^