[Question] CodeAct Architectures

[austinmw]

Hi, I'm curious how you see CodeAct architectures fitting into MCP? CodeAct has seen some recent buzz with the introduction of Manus and other chat agents which use it and demonstrating improved tool calling speed and accuracy.

Right now LLMs are using tools as black boxes through API supported tool-use via writing JSON tool calls.

If CodeAct becomes more performant and we transition to importing code modules as tools and executing them in sandboxes, a standard will still be required for defining these modules in the best way for LLMs, and we'll also need a way to serve access to these code module tools.

Curious if there are any thoughts that can be shared around this?

[sohamganatra]

Hey @austinmw, So we are very bullish on "Code Generation for Function Calling" as we have observed some very good early results testing it. I believe there are MCP compatible ways to do it and I am planning to create a small prototype soon to test it out.

I agree a new standard might emerge for CodeAct but MCP can also be made compatible with few tweaks.

[austinmw]

Appreciate your response, would love to see a prototype when you're ready to share!