Agents Are Making Top-Down Software Development Real

Traditionally, software was built from the ground up: one line of code, one file, one feature, one test, one system boundary at a time. That work still matters. It is how software becomes reliable, understandable, maintainable, and safe.

But agents make another mode practical too: a top-down approach.

I call this "code carving."

You can start with a detailed plan and, using good grounding and tight agent-looping protocols, have agents turn that plan into a rough working application or feature. Then you can inspect the result and carve out what you want from it.

Historically, the "bottom up" approach depended on heavy planning: PRDs, Figma mockups, prototypes, specs. It made sense. Planning was much cheaper than building, so teams had to be risk-averse.

But a plan is not a product. A plan is an educated guess at what the first draft of the product should be. Real product development begins when people start using your product.

Now we can speedrun the first leg of the race: get something out there that people, including yourself, can actually use. And editing the thing, steering it this way or that, is no longer a Herculean effort.

Agents unlock the "top-down" approach to building: instead of starting with a blank canvas and meticulously applying strokes, one blob of paint at a time, you can start with a block of marble and chisel away at what doesn't make sense, revealing the human form inside.

That is code carving: starting with a detailed plan, having agents turn it into rough working software, then shaping the result by judgment, pruning, and proof.