agents write code 10x faster but debugging is the same speed it was in 2019

Posted by notomarsol@reddit | ExperiencedDevs | View on Reddit | 13 comments

something broke in prod yesterday. by the time i opened slack there were already messages about it, and 4 PRs from the last day that could be the cause, all of them passed CI, all of them got approved, all of them looked clean when i reviewed them.

it took me 3 hours to find it. race condition in a queue handler that only shows up when two events fire within about 40ms of each other.

the agent that wrote the code couldn't help me figure out why it broke. i tried, multiple times, different prompts. the bug wasn't in the code it generated, it was in how that code talked to another service from 8 months ago that someone who already left wrote. you cant prompt your way out of "this requires understanding the full system."

and this is where i'm losing it a little. the dashboards say we"re shipping 2-3x more PRs per week now. our leadership keeps sending around the metrics. but the time i actually spend AT MY DESK has barely changed, it just got rearranged. i write less code, i sit in more incident reviews

we have all the AI stuff btw. Claude Code, Cursor, CodeRabbit, Sentry, Datadog, the usual. coderabbit catches the obvious stuff at PR time which is great because it means the bugs that DO ship are the weird ones, the system-interaction stuff, the race conditions, the cross-service timing things. that part is still me at 2am with coffee, scrolling through traces

ok so the post-incident work has gotten HARDER not easier. when juniors wrote bad code i could usually figure out their reasoning by reading the code, even if the reasoning was wrong it was there. when an agent writes bad code there's no reasoning to recover. it just pattern-matched something that almost worked

different shape of problem entirely. yeah idk. our speed is up, defect rate at first review is down. but the average time it takes me to resolve incidents has gone up and nobody wants to talk about that one because its not on the dashboard

[-]

experienceddevsb@reddit

This flair is only allowed on wednesday, saturday (UTC). Please repost on an allowed day. Intentionally trying to circumvent this rule will result in a suspension. See: https://www.reddit.com/r/ExperiencedDevs/comments/1rfhdrg/moderation_changes/

throwaway_0x90@reddit

Hmmmm... race conditions are hard to find regardless if the code is easy or hard to understand.

"but debugging is the same speed it was in 2019"

Hmmm... not necessarily.

Flashy-Whereas-3234@reddit

I agree, but

Just because we can pump out more code, doesn't mean we should still be pumping out the same dogshit broken windows traditional code.

If you're going to really harness this power, then having more "complex but simple" systems is a huge benefit. Like, don't just return true/false for permissions, return the reasons. That takes more complex code, but makes things simpler, get it? It's a whole new way of working, and just going faster is flinging shit at the wall

Doing things better requires smart people who don't just enjoy rolling around in magic numbers doing one-shot Yolo work, and Leaders who say "yes it's functional but also fuck you for shifting the burden to operations"

Agents themselves are very shit about reasoning around the passage of time, so your race condition would easily go under the radar unless there's a ton of verbiage around how those queues and processes work, or little tell-tale checks related to the problem.

This also makes agents want to be super verbose in code and docs, which is some of the most disgusting content I've ever seen, polluting even the worst code based with even more trash that doesn't deserve to exist. Docs and comments need to be curated for both the AI and the human, and AI does an awful terrible job that often misleads itself due to being overly specific.

We should be doing the same amount, just far better, instead we're doing way more, way shitter.

itix@reddit

In this case, it was neither AI nor juniors who were the problem. It was that dude who wrote code 8 months ago and left.

AssignmentDull5197@reddit

This matches my experience. Agents raise throughput, but incident resolution still needs deep system context and causal tracing. Maybe the next win is better agent runbooks and telemetry integration, not more codegen. Ive read similar takes at https://medium.com/conversational-ai-weekly.

bystanderInnen@reddit

Skill issue, claude goes brrrr

eoz@reddit

no way, it's as though the job of a software engineer is to develop a shared understanding of a system and to keep it understandable and debuggable and working by appending the mathematical average of every stack overflow answer related to your current JIRA ticket to the codebase repeatedly for months doesn't achieve that

notomarsol@reddit (OP)

I will also buy it from him and not OP.