“Our systematic study exposes a phenomenon of constraint decay in LLM-based coding agents. While current models excel at unconstrained generation, their performance drops when forced to navigate explicit architectural rules. For end-users, this dichotomy implies that agents are reliable for rapid prototyping but remain unreliable for production-grade backend development.”
One major weakness of this study is that they didn’t fully test frontier models for cost reasons, so the specific performance results should be taken with a grain of salt. But the overall conclusion that models degrade when both behavior and architecture must be correct is interesting, and something to keep an eye on.
I think it's downstream of "you can't optimize for two different objectives".
If you only have functional requirements, then in effect you're doing some form of program synthesis, and RL can optimize that very hard.
If you have a mixture of functional and non-functional requirements, you are basically giving the model an incomplete specification, and it must in some way guess at the user's intent to fill in the blanks. This is also why adding to the prompt examples of the style of code you want (hats off to antirez for this particular tip ;)) is phenomenally powerful.
I am obviously paraphrasing, but the general idea is that trying to synthesize style from a codebase into e.g. a markdown guide generally doesn't work very well. What achieves style transfer is providing the model with a lot of examples of the style, conventions, patterns you want.
To put it in practice: if you point claude/codex to a repository and you ask it to implement feature X using style guide Y, the code will probably work, but you can usually get better results by saying "do it in the style of this file, it was done well there".
Even the strongest frontier model they used - GPT 5.2 - I would consider barely usable for agentic programming.
I’m not really interested in analysis of the weaknesses of such models because in my experience many weaknesses disappear entirely as models get stronger and reasoning effort is turned up. Especially if you tell them what you want them to do.
Also, it’s not surprising to learn that when more acceptance criteria are added the failure rate increases.
Reminds me of the recent paper about delegating document editing tasks to LLMs across different disciplines [1]. That paper found that programming was the only discipline most LLMs can perform long horizon tasks on without accumulating errors & corrupting the document.
I've only read the abstract of this one so far but it seems like this paper has zoomed in on programming with greater fidelity and shown a similar phenomenon. But not about long horizon tasks, more like "long style horizons" of larger sets of structural constraints.
I think that’s mostly because they get so much more of that reinforcement learning - since it is so economical. I dont know if there is any evidence of a fundamental reason they can’t be just as good at other tasks, but it might be economically infeasible for awhile yet.
This sounds like another version of "As a chat becomes longer, the guardrails seem to become fuzzy". You can't use all of the context window bc at the end, the output would not respect the constraints (or guardrails) but to reliably produce production grade code you want the model to have expansive awareness which fills up the context window pretty quickly. It's like saying "Keep everything in mind from these 6 directories - and make this <insert ticket> change" - but keeping everything in mind already fills it's context window which makes it lose it's ability to follow the constraints (or guardrails).
> Our findings reveal a phenomenon of constraint decay: as structural requirements accumulate, agent performance exhibits a substantial decline.
I have exactly the inverse findings on my end. The bigger and more legacy the codebase, the more accurate the patches become.
The harness itself seems to be the most important part. I use a recursive loop that primes the root context based on the user prompt each time. My agent will often make over 100 tool calls to sql and git before it finally decides to apply a patch. If I was greenfield, there would be nothing to query or constrain against.
The harness mattering more than the model lines up with my experience too. What this paper measures is within-turn constraint decay. The version that bites in multi-agent setups is across-session — the architectural rules an agent wrote down on Monday don't reach the agent making the next change on Tuesday.
Also they used languages with dynamic typing like Python & JS. In my experience a statically typed codebase is easier to maintain for humans so maybe it is also for agents.
When using Codex/Claude Code with Go code I cannot count the times the agent does some change, runs a build to check for errors, find some and fix them.
It's crazy to me that people think of Python as dynamically typed by default. Strong static typing has been an option in Python for years now, and it should just be the default.
The python type hints are useful for static analysis (and yes, should be the default) but it’s a joke compared to the utility of types in a language like Haskell.
This is interesting, anecdotally I have felt like I was having better luck with raw sqlite than using an ORM in a recent typescript project, using raw sqlite queries vs drizzle
To this day frontier models think that A and not B means A and B when the sentence gets pushed far enough back in their context window. The context length that model can reason over without obvious errors is much smaller than the advertised context. Between a 1/4th to a 1/20th what is advertised on the tin.
One major weakness of this study is that they didn’t fully test frontier models for cost reasons, so the specific performance results should be taken with a grain of salt. But the overall conclusion that models degrade when both behavior and architecture must be correct is interesting, and something to keep an eye on.
If you only have functional requirements, then in effect you're doing some form of program synthesis, and RL can optimize that very hard.
If you have a mixture of functional and non-functional requirements, you are basically giving the model an incomplete specification, and it must in some way guess at the user's intent to fill in the blanks. This is also why adding to the prompt examples of the style of code you want (hats off to antirez for this particular tip ;)) is phenomenally powerful.
To put it in practice: if you point claude/codex to a repository and you ask it to implement feature X using style guide Y, the code will probably work, but you can usually get better results by saying "do it in the style of this file, it was done well there".
I’m not really interested in analysis of the weaknesses of such models because in my experience many weaknesses disappear entirely as models get stronger and reasoning effort is turned up. Especially if you tell them what you want them to do.
Also, it’s not surprising to learn that when more acceptance criteria are added the failure rate increases.
I've only read the abstract of this one so far but it seems like this paper has zoomed in on programming with greater fidelity and shown a similar phenomenon. But not about long horizon tasks, more like "long style horizons" of larger sets of structural constraints.
[1] https://arxiv.org/abs/2604.15597
Discussion: https://news.ycombinator.com/item?id=48073246
I have exactly the inverse findings on my end. The bigger and more legacy the codebase, the more accurate the patches become.
The harness itself seems to be the most important part. I use a recursive loop that primes the root context based on the user prompt each time. My agent will often make over 100 tool calls to sql and git before it finally decides to apply a patch. If I was greenfield, there would be nothing to query or constrain against.
I usually find I can achieve 90% of the outcome I'm trying to achieve. I use sonnet for planning, qwen for coding, sonnet for review.
When using Codex/Claude Code with Go code I cannot count the times the agent does some change, runs a build to check for errors, find some and fix them.