The quality of AI-assisted software depends on unit of work management

most SWE folks still have no idea how big the difference is between the coding agents they tried a year ago and declared as useless and chatgpt 5 paired with Codex or Cursor today

thanks for the article, it's a good one

It's true that I haven't been a hardcore agent-army vibe coder, I just try the popular ones once in a while in a naive way (isn't it the point of these tools, to have little friction ?), claude code for example. And it's cool ! But imperfect, and as this article attests, there's a lot of mental overhead to even have a shot at getting a decent output. And even if it's decent, it still needs to be reviewed and could include logical flaws.

I'd rather use it the other way, I'm the one in charge, and the AI reviews any logical flaw or things that I would have missed. I don't even have to think about context window since it'll only look at my new code logic.

So yeah, 3 years after the first ChatGPT and Copilot, I don't feel huge changes regarding "automated" AI programming, and I don't have any AI tool in my IDE, I pefer to have a chat using their website, to brainstorm, or occasionally find a solution to something I'm stuck on.

> most SWE folks still have no idea how big the difference is between the coding agents they tried a year ago and declared as useless and chatgpt 5 paired with Codex or Cursor today

yes, just as was said each and every previous time OpenAI/anthropic shit out a new model

"now it doesn't suck!"

I use agents for coding small stuff at work almost every day. I would say there has been some improvement compared to a year ago but it’s not any sort of step change. They still are only able to complete simple “intern-level” tasks around 50% of the time. Which is helpful but not revolutionary.

I think most SWEs do have a good idea where I work.

They know that its a significant, but not revolutionary improvement.

If you supervise and manage your agents closely on well scoped (small) tasks they are pretty handy.

If you need a prototype and don't care about code quality or maintenance, they are great.

Anyone claiming 2x, 5x, 10x etc is absolutely kidding themselves for any non-trivial software.

Last week I wanted to generate some test data for some unit tests for a certain function in a C codebase. It's an audio codec library, so I could have modified the function to dump its inputs to disk and then run the library on any audio file and then hardcoded the input into the unit tests. Instead, I decided I wanted to save a few bytes and wanted to look at generating dummy data dynamically. I wanted to try out Claude for generating the code that would generate the data, so to keep the context manageable I extracted the function and all its dependencies into a self-contained C program (less than 200 lines altogether) and asked it to write a function that would generate dummy data, in C.

Impressively, it recognized the structure of the code and correctly identified it as a component of an audio codec library, and provided a reasonably complete description of many minute details specific to this codec and the work that the function was doing.

Rather less impressively, it decided to ignore my request and write a function that used C++ features throughout, such as type inference and lambdas, or should I say "lambdas" because it was actually just a function-defined-within-a-function that tried to access and mutate variables outside of its own function scope, like we were writing Javascript or something. Even apart from that, the code was rife with the sorts of warnings that even a default invocation of gcc would flag.

I can see why people would be wowed by this on its face. I wouldn't expect any average developer to have such a depth of knowledge and breadth of pattern-matching ability to be able to identify the specific task that this specific function in this specific audio codec was performing.

At the same time, this is clearly not a tool that's suitable for letting loose on a codebase without EXTREME supervision. This was a fresh session (no prior context to confuse it) using a tightly crafted prompt (a small, self-contained C program doing one thing) with a clear goal, and it still required constant handholding.

At the end of the day, I got the code working by editing it manually, but in an honest retrospective I would have to admit that the overall process actually didn't save me any time at all.

Ironically, despite how they're sold, these tools are infinitely better at going from code to English than going the other way around.

I'd argue this just proves my point.

Each and every new model expands the scope of what you can do. You notice that, get elated when things that didn’t work start working, then three weeks later the honeymoon period is over and you notice the remaining limits.

The hedonic treadmill ensures it feels the same way each time.

But that doesn’t mean the models aren’t improving, nor that the scope isn’t expanding. If you compare today’s tools to those a year ago, the difference is stark.

I feel this. I've had a few tasks now where in honest retrospect I find myself asking "did that really speed me up". Its a bit demoralising cause not only do you waste time, you have a worse mental model of the resulting code and feel less sense of ownership over the result.

Brainstorming, ideation and small, well defined tasks where I can quickly vet the solution : these feel like the sweet spot for current frontier model capabilities.

(Unless you are pumping out some sloppy React SPA that you don't care about anything except get it working as fast as possible - fine, get Claude code to one shot it)

There’s been a lot of noise about Claude performance degradation, and the current best option is probably Codex, but this still surprises me. It sounds like it succeeded on the hard part, then stumbled on the easy bit.

Just two questions, if you don’t mind satisfying my curiosity.

- Did you tell it to write C? Or better yet, what was the prompt? You can use Claude --resume to easily find that.

- Which model? (Sooner or Opus)? Though I’d have expected either one to work.

I still use Claude Code and Cursor and tbh still run into a lot of the same issues. Hallucinating code, hallucinating requirements, even when scoped to a very simple "make this small change".

It's good enough that it helps, particularly in areas or languages that I'm unfamiliar with. But I'm constantly fighting with it.

I tried again recently and I see absolutely no difference. If there's been some improvement, it's very subtle.

There's a big difference with their benchmarks and real world coding.

Sooner -> Sonnet

She is choosing GPT5 as the good example? Maybe Claude, maybe..

> If you supervise and manage your agents closely on well scoped (small) tasks they are pretty handy

Compared to just doing it yourself though?

Imagine having to micromanage a junior developer like this to get good results

Ridiculous tbh

if the benefit is less than 2x then we're talking about AI assisted coding as being a very, very expensive IntelliSense. 1.x improvement just isn't much. My mind goes back to that study showing engineers claimed a 20% improvement and measured 20% reduction in productivity -- this is all encouraging me to just keep using traditional tools.

I've found it to be insanely productive when doing framework-based web development (currently working with Django), I would say it's an easy 5-10x improvement in productivity there, but I still need to keep a somewhat close eye on it. It's not nearly as productive in my home grown stuff, it can be kind of annoying actually.

> Ironically, despite how they're sold, these tools are infinitely better at going from code to English than going the other way around.

Yes. Decently useful (and reasonably safe) to red team yourself with. But extremely easy to red queen yourself otherwise.

I've found a pretty good speed up just letting Claude Code run with a custom prompt to gather the context (relevant files, types, etc..) for the task then having it put together a document with that context.

It takes all of five minutes to have it run and at the end I can review it, if it's small ask it to execute, and if it actually requires me to work it myself well now I have a reference with line numbers, some comments on how the system appears to work, what the intent is, areas of interest, etc..

I also rely heavily on the sequential thinking MCP server to give it more structure.

Edit:

I will say because I think it's important I've been a senior dev for a while now, a lot of my job _is_ reviewing other people's pull requests. I don't find it hard or tedious at all.

Honestly it's a lot easier to review a few small "PRs" as the agent works than some of the giant PRs I'd get from team members before.

> I've been a senior dev for a while now, a lot of my job _is_ reviewing other people's pull requests

I kind of hate that I'm saying this, but I'm sort of similar and one thing I really like is having zero guilt about trashing the LLM's code. So often people are submitting something and the code is OK but just pervasively not quite how I like it. Some staff will engage in micro arguments about things rather than just doing them how I want and it's just tiring. Then LLMs are really good at explaining why they did stuff (or simulating that) as well. LLMs will enthusiastically redo something and then help adjust their own AGENTS.md file to align better in the future.

The only AI-assisted software work I've seen actually have a benefit is the way my coworker use Supermaven, where it's basically Intellisense but suggesting filling in the function parameters for you as well. He'll type `MergeEx` and it will not just suggest `MergeExample(` as Intellisense would have done, but also suggest `MergeExample(oldExample, newExample, mergeOptions)` based on the variable names in scope at the moment and which ones line up with the types. Then he presses Tab and moves on, saving 10-15 seconds of typing. Repeat that multiple times through the day and it might be a 10% improvement, with no time lost on fiddling with prompts to get the AI to correct its mistakes. (Here, if the suggestion is wrong, he just ignores it and keeps typing, and the second he types a character that wasn't the next one in the suggestion it goes away and a new suggestion might be calculated, but the cognitive load in ignoring the incorrect suggestion is minimal).