[FEAT]: Support for Multi-Agent Functionality

[NYKD-54]

What would you like to see?

Description

The trend in AI is increasingly moving towards multi-agent support, with platforms like ChatGPT and LangChain already providing robust multi-agent functionalities. This allows for more dynamic and complex workflows by enabling agents to collaborate on tasks, divide work, and enhance overall efficiency.

From what I understand, AnythingLLM currently only supports single-tool calling within its AI Agent functionality. Expanding to multi-agent support could potentially enable more sophisticated task management and cross-functional capabilities within the framework.

Question

Given the current architecture of AnythingLLM, is it possible to implement a multi-agent framework? If not, is there a roadmap outlining plans to introduce this functionality in the future? This feature could further enhance the versatility of AnythingLLM and expand its use cases significantly.

 

Thank you for the incredible work on this platform—it’s a powerful tool that keeps advancing in valuable ways!

[timothycarambat]

Given our arch we can easily (and some custom forks already do) support multi-agent arch. We use heavily modified build of aibitat

So this kind of details:
https://github.com/Mintplex-Labs/anything-llm/blob/da3d0283ffee9c592e5b81d2be6a848722df298f/server/utils/agents/aibitat/example/websocket/websock-multi-turn-chat.js#L77C5-L84C8

Could be done in invocation of an agent session - it just has not been supported because it could get messy quick

[NYKD-54]

Thank you!!
I have another question:

What if we want to use sequence or parallel tool/function calling?
Like the example below

resouce: gemini docs

I try the prompt "What is the weather in Tokyo and London?", but didn't trigger two calling.
I'm not sure whether to implement this capability from the model or the application?