Fundamentally, steering LLMs requires the same structured, logical thought process that is required to write code, regardless of abstraction level. Unlike what HN would have you believe this is not a skill that is equally distributed across the population.

But given the rapid pace at which this technology is evolving, "steering" may very well be ceded to the clankers. LLM agents are fantastic at logical reasoning & any inefficiencies relative to human experts can be circumvented by sheer compute.