A powerful thing for LLMs is this means (I believe!) the metaobject protocol around most functionality -- introspection/reflection/dependency injection -- can b

A powerful thing for LLMs is this means (I believe!) the metaobject protocol around most functionality -- introspection/reflection/dependency injection -- can be made quite powerful. So then all of a sudden, the LLM can act on everything. E.g., today you can ask chatgpt to run code in a sandbox, but not run code to edit your chat session. A more powerful mode would be the functional equivalent: let the LLM not only simulate acting on everything in just the neural model, but generate a copy for agentic models to test manipulating them in reality, and backtrack as needed.