Even if the ReAct paper was published in 2019, I don't think GPT-2 was robust enough to actually work with a tool-calling approach even when finetuned.

For regular coding, GPT-2 was effectively useless because it was only trained from links posted on Reddit.