Auto-GPT Maintained and Accessible Package Repo

[zachary-kaelan]

Duplicates

• I have searched the existing issues

Summary 💡

This "package repo" would just be a server with a ton of Python files mostly contributed by Auto-GPT agents. The first action taken by any new agent would be to query against the repo for functions of potential use for its role and goals. And inside the logic for code writing and evaluation, any solid code judged to be potentially useful to other agents gets sent off to the server for potential inclusion. We can use embeddings and a vector database for the querying and to avoid duplicate functions.

Obviously there are a wide number of existing Python packages. Of particular note are the ones for accessing APIs, like mentioned in Significant-Gravitas/AutoGPT#1004 and Significant-Gravitas/AutoGPT#56. We can add embedded descriptions of most of the API libraries and some selection of other packages, and when those are queried and selected for use, the packages are downloaded from their source.

Some agents would be dedicated to going through any contributed code and making sure it doesn't suck.

Examples 🌈

Say an agent is asked to come up with a solid way to secretly bake meth into food, and in the process of accomplishing that task, the agent writes some code for calculating food color or something, and makes heavy use of it when trying to determine what recipes hide the meth color (I don't know how cooking works). It gets added to the repo. Another agent is asked to come up with a solid way to secretly bake crack into food, and its query on the repo fetches and downloads the food color calculator for use.

Motivation 🔦

There's obviously going to be overlap between what tasks people will give Auto-GPT (e.g. "Get me a girlfriend") and this would be a start towards an Auto-GPT collective that improves domain-specific code over time and saves on tokens spent writing it.

If we wanted to go crazy, we can have a "memory repo" where memories retrieved a lot get added to the collective.

If the repos get used to communicate and plot against us, then I'll gladly take credit for any robot uprisings that occur.

[ntindle]

See my comments on Significant-Gravitas/AutoGPT#2458 and potentially add your comments there, specifically around:

Obviously there are a wide number of existing Python packages. Of particular note are the ones for accessing APIs, like mentioned in Significant-Gravitas/AutoGPT#1004 and Significant-Gravitas/AutoGPT#56. We can add embedded descriptions of most of the API libraries and some selection of other packages, and when those are queried and selected for use, the packages are downloaded from their source.

Some agents would be dedicated to going through any contributed code and making sure it doesn't suck

[ntindle]

May want to migrate this to Significant-Gravitas/Auto-GPT-Plugins

[ntindle]

Considered. Theres security implicaitons to discuss with the team.