Comment by ilaksh

This is a side issue, but he mentions that the harness is stateless. But no agentic task is actually stateless.

I suspect for purposes of optimization and performance etc., sometime in the next several months we will start to see new popular ML architectures or variations of LLMs/VLMs that are designed to be _stateful_. So a lot of the infrastructure for managing state and memory etc. gets sucked into the model somehow.

They may end up changing or expanding the concept of an ML model to enable that.

That type of belief is what is making me anxious about AI engineering staying relevant for much longer. I just wonder if pretty soon we need to be able to build and train or customize ML models that just do everything, or at least know how to prompt an agent to do that.