The bewildered feeling is doing real work, I think. Reading my own code I'm checking it against a design I already have in my head. Reading someone else's I have to rebuild the design first, and that's where most of the bugs turn up.
What surprised me is that writing the ticket myself didn't help much. The agent satisfies it in ways I wouldn't have picked, so I'm still rebuilding.
Author here. This came out of the last couple of weeks at work, where review has crowded out writing code because teammates and agents both produce it faster than anyone reads it. The short version: review the approach before the lines, and give AI-generated changes less benefit of the doubt, since fluency no longer signals correctness.
The outcome might not be deterministic (edge-case, race condition, weird error path). This is also true with human but at least a decent engineer can sit in a review board walking through their reasoning on why did they do what they did and their rationale
Good luck asking AI on "why use X, why X and not Y", yes they will answer but it is based on their current, ad-hoc deductions, not their past reasoning
I guess the same can be said for legacy code base where the original author had left years ago and currently running a kebab shop in Dagestan but my point still remains
Commits, has a PR thread, maybe a ticket. There's something to dig into, and an agent is decent at excavating it.
Ask the agent why X and not Y about its own code and you get a fresh answer made up on the spot with no connection to what happened when it was written.
I'd try Qwen 3.6 35B-A3B or 27B with a temperature of around 0.6. You'd need to offload some of it to RAM though. I've had okay luck with Gemma 4 12B with a temperature of 1.0 but it will miss things a lot :(
Thank you. Gemma4:e4b with temperature 0.9 works pretty good with tools (Bash, Glob, Red/Write and WebFetch) And it is fast. Will try the Qwen 3.6 later today.
