execute agent generated code

[busssard]

https://github.com/KillianLucas/open-interpreter

The interpreter will execute code that has been generated by the agent.
I feel like for a perfect swarm there is the requirement of the individual agents being locally run.

There is a new paper where one can run multiple finetuned instances of an LLM for free, which would allow multiple agents to run with the cost of one.
Then if we have two or three of these local models one would be able to improve the local code of one model and let the swarm get to work improving itself. https://github.com/punica-ai/punica

[caillef]

I like how interpreter and aider can do actions on the system and use local files in a single command. Could be cool to have a demo of several interpreter talking to each other to see how powerful they can be.

[streichsbaer]

I also like the pattern provided by AutoGen, check out their FAQ on Code Execution.

There is one dedicated agent (User Proxy Agent) that does the code execution. That agent receives the code to be executed that was generated by other agents. You can define if it runs locally on the system, or if it's sandboxed to a Docker container.