Ask HN: Is anyone experimenting with running an LLM agent indefinitely?
I'm used to seeing agents woken up to handle a specific tasks, which they perform and then halt - but I'm interested in a situation of an agent given full control of a VM and being run indefinitely (e.g. continuously triggered with the same prompt, or by two sub-LLMs driving each other).
For example, I'm thinking about having an LLM developer with the general goal of improving a piece of software given a general roadmap and some weighted metrics to optimize (e.g. keep test errors at zero, keep code complexity low, keep resource usage low costs, keep runtime latency low). I'd love to see what such software would evolve to. Or equivalently, what it would it look like if we asked an LLM to write and continuously rewrite/edit a short story - would it converge toward something that's worth reading?
Anyone working on this sort of stuff?