The era of intelligent agents is in full swing, and a plethora of new frameworks are emerging. When developers build intelligent agents, they need to choose the right development framework. This article compares three intelligent agent development methods: pure code without using a framework, LangGraph, and LlamaIndex Workflows.
The test intelligent agent has multiple functions, including executing function calls, using various tools, connecting to external resources, and sharing state memory. It is built in the form of a chatbot. The user interface is built with Gradio and has the capabilities of answering questions, interacting with data, and providing insights. It needs to master skills such as RAG combined with documents, generating SQL statements, and data analysis.
The pure code method is simple and direct, and the code structure is clear. It is a good way to learn the principles of intelligent agents. However, it is difficult to design the system prompt words for the skill router, and handling different output formats is complex.
LangGraph has a long history and was released in January 2024. It is built on LangChain and uses a Pregel graph structure to simplify the creation of cyclic processes in intelligent agents. The concepts of nodes, edges, and conditional edges make graph traversal more intuitive. Its architecture is similar to that of a pure code intelligent agent, but the underlying code and the mechanism for controlling skill switching are different. When using LangGraph, one may face function call validation errors and difficulties in debugging. However, it is easy to use, and the graph structure code is concise and easy to understand. It can also convert existing applications built with LangChain.
LlamaIndex Workflows is a new member of the intelligent agent framework that debuted in early summer this year. Similar to LangGraph, it aims to simplify the construction of recyclable intelligent agents and emphasizes asynchronous execution capabilities. Its architecture is similar to LangGraph but the code implementation is different. It adds an initialization step to prepare the context of the intelligent agent environment. When using Workflows, one may face difficulties in synchronous-to-asynchronous conversion and Pydantic Validation Errors. However, the advantage is that one needs to write their own routing logic and tool operation code, similar to pure code intelligent agents. The event-based architecture is helpful for managing complex asynchronous systems. It is lightweight and has fewer restrictions.
Comparing the three methods, the no-framework method is simple to implement. Its readability and ease of use depend on the developer and may become difficult to manage when it becomes complex. LangGraph provides rich structural support, which is beneficial for multi-person collaboration and beginners, but if one does not agree with its framework, it may be a headache. Workflows is somewhere in between. The event-based architecture is valuable, and it has less strict requirements on the LlamaIndex type and provides greater freedom.
When choosing an intelligent agent framework, three key questions need to be considered: whether the project has deeply integrated LlamaIndex or LangChain; whether one is familiar with the intelligent agent architecture and needs guidance; and whether there are reference examples. Choosing an intelligent agent framework is only one of the decisions for the production environment of generative AI systems. We also need to establish security safeguards and monitor large language models and remain flexible in response.
