[Question]: Information on Agents

[sreenivasanm6]

Question Validation

• [X] I have searched both the documentation and discord for an answer.

Question

Why should one prefer llama index for agents when compared to langgraph? Please help in providing detail.

[dosubot[bot]]

I wasn't able to find specific information comparing LlamaIndex and LangGraph for agents in the repository. However, LlamaIndex offers a comprehensive framework for building agentic systems with features like breaking down complex questions, choosing external tools, planning and executing tasks, and storing completed tasks in a memory module. It supports various use cases, provides prebuilt agent architectures, and allows for custom workflows. Additionally, LlamaIndex can be integrated as a tool within other agent frameworks, enhancing its versatility ^[1][2].

To continue talking to Dosu, mention @dosu.

[sreenivasanm6]

@dosu not clear in specific to langgraph

[logan-markewich]

@sreenivasanm6 any agentic implementation in llama-index should probably be using workflows

While we do have prebuilts like FunctionCallingAgent to get started quickly, in most cases I see, users need so much customization that it makes sense to build more from-scratch with workflows Workflow Docs Function Calling Agent w/ Streaming, Dynamic Context

I can say that in my own observations/notes

• Workflows will likely be cleaner, easier to understand, and less LOC than anything you can build with langgraph (there's a reason they've prioritized building stuff like langgraph studio)
• Workflows, in comparison to other frameworks, are async first, which is super valuable in production workloads
• The event-driven approach used by Workflows allows users to focus on what data each component requires in order to run and what data should be sent back, without the need of detailing the execution flow upfront.
• Workflows naturally allow loops and arbitrary jumps from one component to another during execution.
• Steps are in full control of the data they receive, making it easy to manage optional and default values.

The feedback on workflows has been positive from basically everyone who's tried it. I would say that the only negative that comes to mind is that not every developer is familiar with async+python

[sreenivasanm6]

Thanks for the info. How about the memory consumption, token limit usage of api calls etc. Will it be less or more when compared to langraph?