Meanwhile I'm over here refactoring as much as I can from years (or decades) of human-slung code. Turning the mess I either inherited, helped create, or built on top of into something clean and pristine might be my favorite LLM use. Same for personal projects, codebases that evolved over many years when I happened to have time that weren't kept quite as "clean" as I wish that finally been cleaned up.
I've always _wanted_ my code to be clean and easy to follow but life, deadlines, shifting-priorities, etc have stood in the way of that. Now I can finally realize my personal nirvana.
That said, I've had to steer models away from too-heavy of abstraction or similar because it made the code too hard to follow.
A lot of this feels like it comes down to the training of the agents to produce code that satisfies the various benchmarks combined with reactions to things which were previously maladaptive. I.e. things which were explicitly trained out of the model in post training. I think there's a lot of missing long term software engineering principles that don't seem to be baked into the way the models tend to write code by default.
I have some speculation that maybe the people doing the model post-training tend to be younger researchers that haven't worked on large complex software systems, so their taste isn't as developed in this regard about what things are important here.
But this is an area that can be steered with appropriate early instructions ("When choosing tradeoffs of implementation, build for long term maintainability and understandability of code over implementing just the exact code necessary to solve the issues. etc. chain of thought often includes information that would have to be repeated in a future agent session, make sure to persist it to code or external docs so that future sessions and user understanding is respected.")
It can also be done as a post-change step with similar effects. And you can use your agents to build this layer into your general modus operandi for dealing with the crimes of generated code. But one of the things that all AI labs should be doing is looking at AGENTS.md on real project as being hard expressions of what failure modes real projects have noticed in models generally. Don't wait fo the bugs to be raised on these things, use express preferences that show that there's a problem. Go trawl github for these in bulk to use for future post-training.
I've found that access to coding agents has helped me be far less tolerant of bad code patterns that can be refactored.
Refactoring used to have a very real cost - it was substantial amounts of time that would have to be carved away from working on new features.
Now I can spot a potential refactor, fire off a prompt in an asynchronous coding agent (or on a worktree or whatever), then come back 20 minutes later and either accept it, poke it a bit, or abandon it. Costs me almost nothing.
And for me, it has created a giant pile of unmaintainable code from my coworkers because of stupid management people that think they can code.
Refactoring was never a substantial amounts of time for me before llms.
Before I could spot a potential refactor and refactor it in 20 minutes and less. Never abandon it. Just constant improvement to the point the previous tech startup that I was working for just drive from itself (I am still paid a a few hours per months for it)
If you were fortunate enough to be working in a project with very good test coverage, refactors were easier (maybe not easy) in the sense that you were at least fairly confident that nothing broke.
But I've personally never worked on code with test coverage that good. Refactors were always risky.
Do you work with other, potentially unmotivated devs/managers a lot? I wonder how much of this is an incentive issue. I don’t think refactoring itself is as much of an issue than working with 2-3 other engineers that DGAF and a manager who is only looking at LOC to determine who to promote.
But it also doesn’t mean these things aren’t problems, they’re obviously huge problems.
All sorts of efforts that used to be put off indefinitely can now be handled largely by LLMs. Yesterday I took a codebase (~50 source files) and spawned a sub-agent (GLM 5.3 Flash) for every single file. Each file was analyzed for test coverage issues, inconsistencies between comments and implementations, and all call sites against the implementation. Then issues were aggregated. I reviewed the list manually, had Opus and K3 review as well to prune the list, and then had a commit made for every minor issue found. 70 commits with maybe 30 minutes of manual work.
Not exactly a refactor, but high degrees of consistency are what I strive for in a codebase. LLMs get confused by inconsistencies, as do humans.
AI agents get lost all the time, particularly if the codebase is already sprawling out of control.
Your discipline only pays off if you already understand your code and/or established clear baseline for your standards before launching into a feature development mania. And it needs to be enforced every turn, or the firehose of code generation knocks the front door down easily.
Let's say that tomorrow, due to an improved model or whatever, we realize that the most efficient form of code of an app - for an llm to understand and work with - is for it to be in one long spaghetti file.
Why wouldn't we do that? I think there's a point where this comes down to values instead of facts. If you want it to be human readable, that's fine and there are a bunch of therefores from that point. But if you don't necessarily want that for a particular codebase, why refactor if the LLMs can handle it?
I'd argue the things that make a codebase more human-readable is also what makes it more agent-readable though. Agents are trained on human data, after all.
The position in the article is reasonable because what would end up happening otherwise is:
- Agents increase complexity, humans can't read it anymore
- Agent unable to keep making updates without looping forever (the complexity of the code exceeds the agent's context length). Human doesn't understand either so can't fix.
This isn't hypothetical either, it's basically what ends up happening to most vibe-coded software if the person doing the vibe coding doesn't know how to review the outputs being produced.
I've personally adopted doing multiple refactoring passes after any large code implementation done by AI. In pretty much any scenario where I'm adding code using AI, it's 1 turn to add the feature and and then another 4-5 turns to refactor and clean everything up.
Often times it's not even that the code is bad but rather that it's overengineered. I see it happen so much that I'm tempted to actually go the other way on a toy project. Like what would Claude or Codex come up with if I told it I wanted an enterprise grade, globally scalable, compliant and auditable tic-tac-toe game.
> Here’s the problem. AI agents are not bound by human context limits in the same way. An agent can read the tangled function, trace every caller, and make sense of the mess that would have stopped a human cold. It can add the next branch correctly, and the one after that, working confidently inside code that no human on the team fully understands anymore.
They're not bound by the same limits but they're still bound by some limits, yeah?
I'm not an AI expert, so I don't honestly understand why LLM driven agents are as good as they are. But my impression is "trace every caller", most of the time, is still an approximation. Once the code has gotten convoluted enough, cases are going to get dropped.
They also very often do not actually read the fucking files! Even after specifying to „read in full the fucking files“. LLMs are so lazy, they take any shortcut they can.
Sorry, just thinking about it is reviving my frustration…
This is pretty spot on imo. Although otoh I feel like I don't need to be able to reason as deeply about a system because of the ability of an agent to dig through a code base. Personally I believe I'm still looking for that new balance.
It's a little like driving a car in a neighborhood you know, vs one where you constantly need to be looking down at a map or gps. It's much more comfortable driving around the places you're familiar with, but you can't know everywhere.
I find the basic premise for low level code quality to be true, in my experience, but counterintuitively I can now police the overall structure and system architecture MUCH more heavily.
As ever, no one is willing to allocate time for this, but (with unlimited work tokens) I can parallel path massive cleanup refactors all the time now.
I've had the opposite experience where we're refactoring the gnarliest shit anyone's ever seen because AI can actually understand it well enough to decompose, test, refactor, etc.
I've always _wanted_ my code to be clean and easy to follow but life, deadlines, shifting-priorities, etc have stood in the way of that. Now I can finally realize my personal nirvana.
That said, I've had to steer models away from too-heavy of abstraction or similar because it made the code too hard to follow.
I have some speculation that maybe the people doing the model post-training tend to be younger researchers that haven't worked on large complex software systems, so their taste isn't as developed in this regard about what things are important here.
But this is an area that can be steered with appropriate early instructions ("When choosing tradeoffs of implementation, build for long term maintainability and understandability of code over implementing just the exact code necessary to solve the issues. etc. chain of thought often includes information that would have to be repeated in a future agent session, make sure to persist it to code or external docs so that future sessions and user understanding is respected.")
It can also be done as a post-change step with similar effects. And you can use your agents to build this layer into your general modus operandi for dealing with the crimes of generated code. But one of the things that all AI labs should be doing is looking at AGENTS.md on real project as being hard expressions of what failure modes real projects have noticed in models generally. Don't wait fo the bugs to be raised on these things, use express preferences that show that there's a problem. Go trawl github for these in bulk to use for future post-training.
Refactoring used to have a very real cost - it was substantial amounts of time that would have to be carved away from working on new features.
Now I can spot a potential refactor, fire off a prompt in an asynchronous coding agent (or on a worktree or whatever), then come back 20 minutes later and either accept it, poke it a bit, or abandon it. Costs me almost nothing.
Refactoring was never a substantial amounts of time for me before llms. Before I could spot a potential refactor and refactor it in 20 minutes and less. Never abandon it. Just constant improvement to the point the previous tech startup that I was working for just drive from itself (I am still paid a a few hours per months for it)
In the only scenario where that could actually be true you wouldn't have coworkers creating unmaintainable code with LLMs now.
But I've personally never worked on code with test coverage that good. Refactors were always risky.
But it also doesn’t mean these things aren’t problems, they’re obviously huge problems.
Not exactly a refactor, but high degrees of consistency are what I strive for in a codebase. LLMs get confused by inconsistencies, as do humans.
Your discipline only pays off if you already understand your code and/or established clear baseline for your standards before launching into a feature development mania. And it needs to be enforced every turn, or the firehose of code generation knocks the front door down easily.
Why wouldn't we do that? I think there's a point where this comes down to values instead of facts. If you want it to be human readable, that's fine and there are a bunch of therefores from that point. But if you don't necessarily want that for a particular codebase, why refactor if the LLMs can handle it?
The position in the article is reasonable because what would end up happening otherwise is:
- Agents increase complexity, humans can't read it anymore
- Agents increase complexity, agent can't read it's own code anymore
- Agent unable to keep making updates without looping forever (the complexity of the code exceeds the agent's context length). Human doesn't understand either so can't fix.
This isn't hypothetical either, it's basically what ends up happening to most vibe-coded software if the person doing the vibe coding doesn't know how to review the outputs being produced.
Often times it's not even that the code is bad but rather that it's overengineered. I see it happen so much that I'm tempted to actually go the other way on a toy project. Like what would Claude or Codex come up with if I told it I wanted an enterprise grade, globally scalable, compliant and auditable tic-tac-toe game.
https://github.com/enterprisequalitycoding/fizzbuzzenterpris...
They're not bound by the same limits but they're still bound by some limits, yeah?
I'm not an AI expert, so I don't honestly understand why LLM driven agents are as good as they are. But my impression is "trace every caller", most of the time, is still an approximation. Once the code has gotten convoluted enough, cases are going to get dropped.
Sorry, just thinking about it is reviving my frustration…
Thats… not my experience. Like, not at all. They very regularly get lost
As ever, no one is willing to allocate time for this, but (with unlimited work tokens) I can parallel path massive cleanup refactors all the time now.
I've yet to work at a place that bothered with much refactoring over adding the thirtieth conditional to new feature....