Quite organically, and unironically, I blurted out the phrase hardcoded program in a conversation this week. It seemed a bit oxymoronic at first, but with reasoning agents getting good enough, a higher level of declarative abstraction is becoming my go-to choice when building AI products.
It's not the easiest thing to admit, but React & HTML are the reason I've been able to pretend to be an engineer for so long. I can tell them "what" I want to paint on the browser - they handle the "how". SQL helps me keep my mask on with databases, and Terraform with infrastructure.
For non-deterministic output, writing imperative code now seems increasingly futile, less reliable and importantly, less dynamic than using a declarative approach. By engineering the right harness, tools and memory - I've been able to get better results by telling an agent the "what" and letting it figure out the exact "how" dynamically. In some scenarios, even for deterministic output, I see this approach already winning over "hardcoded software". (A language model with the right harness is after all Turing Complete). I'm not fully sold that this second set of scenarios will continue to be gobbled up by agents. But if the currently subsidized token economy leads to something more sustainable, then the economics of maintaining hardcoded software will shift dramatically.
For AI engineering though, the shift is already here.
Yet I must keep up my pretense - for now.
First, the judgement, on when to use the declarative approach and when to descend to a lower level, still draws from the fundamentals of an engineering mindset. More importantly, as I let the agent handle the how, it creates the need for, and frees up my time and cognitive capacity for asserting control at a higher level - the what and the why. With traditional declarative programming, I can deterministically verify that the interface reliably produces the "what" that I want it to. With agent engineering, I need a way to do the same - verification & experimentation are core tenets of the scientific method, and reliability that of good engineering and a great product.
In a subsequent post, I will talk about what this experimentation looks like when building agents. This experimentation, after all, is what convinced me of my impending redundancy in the first place.