> are extremely powerful when used with deliberate care. Gas Town is the exact opposite of what is needed to actually do useful things in prod.

This judgement feels premature.

We really don't have any idea what is possible. The wheels within wheels, epicycles within epicycles model of agentic loops hasn't really been deeply explored and we just don't know where they might go.

I too share your instinct that human steering helps, helps a lot. But I've found I can keep using less of it, as I setup better parameters, as I improve conducing the LLM into good paths. The idea that an LLM could sit up top and help try not one idea at a time but try many things, then pick and cobble together next goes: that is madly madly madly exciting to me.

I don't want to keep being the bandwidth limiter in this system: I want to scale out. I haven't been following close or trying but I tend to think while total hands off is not the way, having LLMs that can cover a lot of terrain, explore a lot of solution spaces and directions, then assess how to put it together & what to take forward, and other practices of agents watching agents, has enormous potential.

Deliberate care relies on pre-obtained wisdom, and often, the human biases kind of suck and aren't they good. We aren't great at burning down our systems enough, at Chad Fowler Phoenix Architectures. I think the AI's lack of over deliberation and it's ability to try vastly more could be a huge advantage, could show diversity triumphing over specific crafted intent.