I'm not cutting corners

I enjoyed My week with opencode and the followup The problem is culture by Iris Meredith, an LLM skeptic who gave GLM-4.6 a serious try. I want to dispute a premise stated early in the second post:

cutting corners (which the use of a coding agent fundamentally is and will be for the foreseeable future)

I don’t think this is pedantic because she said “fundamentally is”: this is not what use of a coding agent fundamentally is. To me, what coding agents fundamentally are is a reduction in the cost of producing software and the barrier to getting started (both for the expert and the beginner).¹ What you do with those cost savings is wide open. An experienced software engineer has dozens of ideas in a day that are either immediately rejected or don’t even rise to the level of consciousness because their benefit is not worth the time it would take to execute them. Plenty of those ideas would be worth doing if they were nearly free.

Here are some simple examples of using LLMs to do more rigorous engineering, barely scratching the surface:

• Writing tests that you otherwise wouldn’t have written because they are too tedious. Not just a few extra tests but entire categories of integration test that require a complex setup (examples here of tests that exercise various merge conflict situations in git and jj by spinning up an actual repo and producing the conflicts)
• Formal verification
• Rearranging your changes into more reviewable stacks of pull requests, when before you might have decided it wasn’t worth the effort of teasing your changes apart
• Writing debugging and visualization tools that would otherwise have taken too long to figure out

My colleague Rain and I went on the Oxide podcast last week to discuss how we use LLMs to strengthen rather than avoid rigor in our work. As someone pointed out, this requires “a work culture that prioritizes and enables this level of rigor and care … It’s just not the case everywhere.” I do worry that I’m biased toward optimism because of this, but I think of our culture as an existence proof of something many people seem to think is impossible.²

Another way in which use of agents is fundamentally not about cutting corners is that people who are not software engineers are now developing tools for themselves that would simply not exist otherwise. I’ve seen direct testimony online from journalists, lawyers, policy analysts, and restaurant owners. Obviously they are bringing less rigor to the process than a professional engineer would, but it would be strange to say they’re cutting corners because this software didn’t exist before. Furthermore, this lower level of rigor is appropriate because the tools they’re writing are less complex than most of the software written by professionals, and you can often tell whether it works quite directly. Software is made in a range of conditions, for a range of purposes, to a range of standards of rigor.