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.
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.
Even the best models have trouble adhering to stuff as mundane as rules for how to style generated code (indent this much, name things with these patterns, etc.). Even the most die-hard AI-first coder will admit to that kind of stuff being not unheard-of. Yet they still delude themselves into thinking that these models will follow a sufficiently detailed spec to the letter, every time.
tasks spanning eight web frameworks
The most incredible combo I've seen lately is progressive enhancement of Razor Pages with javascript. With this arrangement the newest models tend to make a really good call on if something should happen server-side (cshtml) or on the client (js).
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.
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".
If there's a second thing the generative AI tools have shown beyond any doubt it's that many of the more modern (relatively speaking) "best practices" that have always been over-hyped and questionably-evidenced really do tend to produce worse results. LLMs take these methods to their logical conclusions and show us the end result much sooner. You can't just iterate your way to a solution when you don't even know what problem you're trying to solve. If you don't have a clear spec then you don't know what a correct product looks like. You need to invest time in reviewing code properly. If you don't keep the big picture in mind then the big picture becomes a mess.
Maybe one day the LLMs will leave me out of a job but at least I'll feel validated first!