I've ended up alternating between "the LLM can write the awful glue code I'm too lazy to hand wring for my proof of concept" and "ooh, what if this was in Fortran _and_ JavaScript?"
replyDark times.
AI workflows don't seem to be suited to image-based environments like Smalltalk's, though.
replyUnless you give them a tool which lets them execute arbitrary code in the running image.
replyOr at least this works wonderfully in Clojure via the nREPL.
On the other hand, ai agent harnesses work a bit like image based system - think: claude --resume
reply