I was a bit hazy on numbers because I don't scientifically record them, but I guess what I'll say is the fixing and verifying takes a lot longer than writing the initial code. This was true before LLMs, but when writing code by hand I had the context of the code in my head, so potential issues, blast radius, etc. was a lot more obvious. By definition most of the things that break are things that are not trivially testable. Unfortunately, it's not as easy as saying "Claude, the thing broke, plz fix"; I've had to spend a lot of time recently helping it with context from debuggers, or just debugging myself manually, adding log statements, etc.

Am I giving the agent enough context? Well, I'm giving it as much as I can. Each submodule has an AGENTS.md, I have the agents add gotchas and instructions for some feature work when I discover where an agent went wrong, the codebase has a lot of comments along the lines of "if you edit this section, you need to also edit XYZ", and I lean on the type system as much as I can to make wrong-code not compile. It has access to playwright for driving the UI if it wants to. (Weirdly, I've found that Claude is really inconsistent about using these tools -- even though the instructions make it clear that it's allowed and encouraged. I think if your workflow differs from the models training and thusly you have to tell it so in AGENTS.md/CLAUDE.md, then it's very inconsistent about following those instructions. For instance, I don't want Claude to commit and I don't want it to sign commit messages, and it still does that all the time even though it's my like #1 directive of "don't mess with my git history")

There are some things though that are very hard for it to test. I'm exporting essentially a programming language to three game engine runtimes. They all have automated tests, but, I think people that have worked in video games know that games are very hard to automate testing on. This isn't really the fault of the agent I would say, just the nature of the problem, but it is worth noting.

I guess this is a long winded way of saying, even with LLMs tech debt is a thing you have to manage, and I think managing tech debt becomes even more important when you're dealing with LLMs, not less important.

"40 hours" in his context here is actually a work day, so 7-8hrs.

He says "40 hours" because he feels like he's managed to do 40 hours worth of work in this time, but then has to spend another "40 hours" (actually: 1 day) just going around kicking tyres.

Obviously the implication is that it's a net gain of some kind, but he's unsure if he caught everything.

(sorry to reiterate the GP, but I feel like you missed the important nuance that it's not a real 40 hours of time).

Wait; So you say you outsource development to a LLM fully knowing that it will not write sound/deterministic code and your quality requirement are your test cases, right?

I further suppose you are not writing the test cases in its whole by hand. But you try to specify them before you hand out the development task to the LLM, right?

It sounds like you just introduces a new team member which is not trustworthy yet. Normally you would review each change of him and explain how to improve hisself and the code. But that's not possible to a fixed-state LLM. That sounds exhausting.

If your Markdown Files should aim at improving the LLM contributions, you are again stuck with the fact that it did not follow in the first place.

So I conclude: You pay for an Intern who is not trustworthy and pay additional input token on Markdown Files to still no be certain about future contributions.

I just don't grasp how we as engineers are accepting this and integrate it into our craft. And: Everybody using External LLM Services, possibly providing the entire project as context, is allowed to let the source code of your company be leaked to some third party. I hope that party is trustworthy and does not have a track record of copyright infridgement. Because this would be fairly naive and reason to be fired. So we as Engineers knowing the implications should therefore point to these issues at the correct management level.

At least that's what I am doing shrugs

I think testing is really important (even without LLMs). Currently I have 2247 unit tests across 132 files in a ~150K LOC codebase (test code included in that count). It takes about 50s to run. There could be more, but it's not nothing. There is playwright for it to test drive the UI, although if I'm being perfectly honest even before LLMs I thought that kind of test tends to be brittle and annoying to write (I guess I don't have to write them anymore, but they are still brittle). I'm honestly trying to give it as much structure as I possibly can -- I'm not trying to setup the agent to fail so I can be like "gotcha!"

I'll also just point out my philosophy for using LLMs for this project, which is that I'm not trying to go as fast as I can. (I want to go at a good pace, but this isn't an experiment to just finish something over a weekend). The 150k LOC have come about since February, with some mix of me writing code and LLMs, so on average I'm probably bringing in about 800 LOC per day, which I imagine a lot of vibers would find to be glacial. To me that's the sustainable rate of what I can do when you factor in that I need to test drive every feature, make sure it doesn't conflict with another feature, check for bugs, check that the code looks reasonable, and debugging. (I also think that rate limit is specific to this project: I could see easier to test things going much faster, and harder to test things going slower)