Reading speed, working memory. Agents are good coders, they are not good middle managers.

For me personally? Local processing power. The reason I rolled my own harness environment is because my 16 core CPU was literally MAXED OUT with parallel compilation units and edit streams. A custom harness barely helped. Just allows me to run a few more sessions in parallel

Should not spending more money on local processing power be a non-brainer then? You will make it all back and then some?

Not 100% related, but I have a draft on a thought experiment: how would change if AI responses were immediate?

Like sub-second latency for _any_ change you ask it. Let it be 100 files changed, MR review and fix of each point, design document, anythikng.

Tool execution (eg: running tests) would be the next tech bottleneck. And human-on-the-loop the obvious one.

Maybe that's where 1000x productivity hides, who knows :)