I wonder if you’ve seen ChatGPT plugins yet. If not, check out the Wolfram blog post.
Just not sure feeding it back to itself does much at the moment.
If it 0-shots something and executes it and halts, not an agent, if it observes output and loops once, that's a minimal agent. I'm not super skeptical of a couple loops; I just don't know that GPT4 has what it takes to drive a larger number of loops.