Autolang as a langchain agent?

[johannhartmann]

Hey,
First: I love autolang, it looks neat and feels like the right method to implement autogpt like services for proprietary tasks.
It just took a few minutes to use it with Llamacpp, and surprisingly, it works quite well - see https://github.com/mayflower/autolang.
However, before supporting multiple LLMs directly in autolang it feels better to use autolang within langchain itself, and your file structure seems a bit like you had the same idea.
Would that be something you are planning to do anyway?

[alvarosevilla95]

Hi, thanks for trying it out and the suggestion!

I am planning to improve the composability with other langchain workflows yeah, but I'm still trying to figure out the right approach. Let me know if you have any ideas in mind!

Right now, each of the system parts is a langchain agent/chain, although I need to make them easier to customize and swap auto in the AutoAgent. Then, more generally, I'm considering the following improvements:

1. Allow multiple Executor agents, the planner is given a description of each and tasks are broken out and assigned to a specific one. So the planner/reviewer become agent orchestrators..

2. Expose autolang itself as a langchain tool. You could then have a layer on top potentially orchestrating / delegating to different systems, maybe customized to more specific workflows.

I think 1. would probably be the right abstraction layer for most uses, so I'll start with that. But adding 2 should be pretty easy.