I think a big part of the problem is the devx / workflows that these tools tend to encourage or impose. Watching my coworkers fumble with terrible copilot autoshit in a pairing session makes me want to scream on a daily basis. That doesn’t necessarily mean such a DX is the only way it could or should work, or that it’s even a good DX at all. Most of the time, it actually imposes more cost, both in lost momentum and in corrective action, than it relieves.

I think something like a graph where nodes are specified clearly with inputs and outputs (or pre- and post-conditions, if you prefer), and agents iteratively generate and refine the implementation details could have some utility, but that would require programmers to know how to think about program design (and have a usable UI/UX for such a tool) instead of hastily stumbling through patches and modifications to existing code they already don’t understand.

The new LLM systems are carrying us farther from that world, not closer to it.