I think the 'novelty' goalpost is being moved here. This notion that agentic LLMs can't handle novel or non-trivial problems needs to die. They don't merely derive solutions from the training data, but synthesize a solution path based on the context that is being built up in the agentic loop. You could make up some obscure DSL whole cloth, that has therefore never been in the training data, feed it the docs and it will happily use it to create output in said DSL.

Also, almost all problems are composite problems where each part is either prior art or in itself somewhat trivial. If you can onboard the LLM onto the problem domain and help it decompose then it can tackle a whole lot more than what it has seen during pre- and post-training.