Code produced by LLMs doesn't get executed properly (empty code and output)

[luijait]

Observing the following issue:

execute_code is a difficult formatting tool for LLMs which leads even the most advanced models to fail when the context window exceeds 30%, there are 2 solutions, delete the tool from a certain context or the best one

As a way to address this issue, we can consider removing the execute_code tool and implement CodeAgent as a tool as proposed by openai-agents-python.

[vmayoral]

This will likely be implemented in 0.4.x, via codeagent Agent