[Jodavaho.io]

I use coding agents daily. So far, they are good for first-order effects. Things like “Fast typewriter” (doing something I would do but without interface delays) or “super copy paste” (altering an example and applying it with nuance in several places) or “super tab complete” (suggesting large blocks of text/code vs just single symbols) or even “meta bash” (asking for scripts that do oneoff analysis or processing of a few data sets you have lying around).

Most of these are code-generation tasks. Sure, we can write code as fast as possible, but that doesn’t help the thinking, decision making, and understanding go faster. According to Amdahls law, that’ll become the blocker, and probably already has. Theoretically, giving up on understanding is the natural next step - essentially let agents build whatever they want. Aka vibe coding.

So we’ve already hit the AI agent first-order improvements limit??

What I want is second order effects. I want tooling around my tooling to help my tooling build my tooling so that I can be more productive and directly involved. That’s the only way to ensure that what I want comes to pass.

I want smart-jump 10x, smart-context popups 10x, live execution / simulation, integrated replay of test cases with backtracking, and a million other things I can’t imagine yet. Instead I get a github agent that automates leaving comments about my code style. Thanks.

Anything else is just unproductive laziness and gambling your future job.