From prompting agents to building them

The way we build software with agents focuses too much on optimizing for efficiency per task rather than rethinking the entire workflow.

Focusing on individual tasks creates busywork that is not worth spending one's time on. Consider reviewing AI-generated code or AI-generated documents that take far less time for the writer to write than for the reader to read.

I see two necessary shifts in how we build software at scale.

For one, engineering time is better spent on building agents instead of business apps. The job of engineers moves from prompting the agents to build the agents. This also means engineers move up the stack - building custom models, designing agent architectures, evaluation loops, and stitching agents into pipelines.

Then a product builder can do the prompting and build features without looking under the hood. We are moving from people having ideas to having full autonomy in implementing them. Think of this role as part product manager, part designer, part business analyst. 

The upcoming challenge for product teams is shifting from making individual tasks faster to reshaping how to build software entirely.