Whilst I understand the intention of code as plan (vs tools) it seems incredibly in-efficient. If you take the customer service agent lab notebook your creating the code each time from scratch through an LLM call for each query if I have understood the code path.
What would be a more practical implementation of this paradigm?
R
Hi jklondon,
Great question!
You can have a look at the paper referenced by Andrew in the video ‘Planning with code execution’ (Wang et al., 2024), published at https://arxiv.org/pdf/2402.01030.
The main use case they mention is model training. Here is a figure from their article:
As Andrew states, the approach can also be effective in building other complex pieces of software.
