The article says what many here like to hear, but in my opinion the core arguments are false.
> Making software debuggable, maintainable, layered, and composable – that’s still quite a trick
Not really. I have been working on a mobile app for months, and I stopped even glancing at the code about two months ago.
150k LOC, around half of that in tests, and the AI still has no problem maintaining the code on my behalf.
Debuggable? It can add extensive instrumentation in seconds.
None of this requires expertise, prompting, or mention of TDD. It's the default.
Frankly I do not believe the author tried developing a large codebase fully agentic and without reviewing the code. I believe many here look at the code produced, deem it substandard, and go hands on.
> They’re foundationally incapable of always and consistently preventing prompt injection attacks
From Anthropic's article about the Auto mode:
> We commissioned an evaluation from a third party, Trajectory Labs, who tested different models within the latest publicly available versions of Claude Code and Codex as of July 17th 2026.1 They tested 72 indirect prompt injection scenarios held out from Anthropic
> In this evaluation, none of the 720 attack attempts succeeded against Claude Fable 5, Opus 5, or Sonnet 5 running auto mode. On the other hand, 5.83% of the attacks succeeded against GPT-5.6 Sol running Codex's Auto-review mode. Notably, this is greater than the 0.09% average attack success rate against our latest models running in bypassPermissions mode without additional safeguards. The tests showed a 19.03% attack success rate against GPT-5.6 Sol when running in Full Access mode
I'm sure someone is going to reply with how they do not trust Antrophic's research, but lacking other data, prompt injection appears to be largely solved already.
mortalapeman 3 hours ago [-]
With generated code, the directory structure, interface design and general state management is usually a haphazard mess. Even with the best frontier models. But what really gets me is the model often tries to make assumptions for me that I didn't specify in the prompt. Subtle things like which error states are "oh shit we need to bail" vs "this isn't a deal breaker." Sometimes it will ask, but more often than not it will just make a decision and it's often the wrong one. If I don't have a fully kitted out test suit and a good type checker to verify the final product against, the the whole looping thing is just useless to me and I'm back to reviewing every line of code it puts out and having to draw on my years of architecture experience to make sure we don't build a giant pile of trash.
Gigachad 2 hours ago [-]
Because they are designed to be used by managers who don't know how to answer these questions and don't want to be asked them. Just have the magic answers box pick something.
bluegatty 1 hours ago [-]
The generated code is fine at the functional level, the directory structure is usually the standard pattern for the given type of project.
The error types and codes, it will produce to spec.
If you type 'make me that thingy' - yes, it's probably not going to do what you want, but if you give it spec and guidance, it usually will.
The 'interface design' ... not very good though.
slopinthebag 2 hours ago [-]
They're RLHF'ed to an inch of their lives to be able to one-shot complete tasks, since requiring human input defeats the purpose of being able to replace the labor force.
But once the insanity ends LLMs will be packaged as tools for developers to use to boost their productivity, and we'll consider them as we do IDE's and debuggers and stuff. But we have to get through this hype cycle first.
dmitrijbelikov 9 minutes ago [-]
LLM is the new Excel
theteapot 2 hours ago [-]
> It helps to know that LLMs don’t “reason”. They predict ..
Semantics. Prediction is the training objective. The ability to reason can be, and very arguably is, an emergent property of that.
complex_pi 25 minutes ago [-]
Maybe it looks like reasoning, and maybe that's enough for some.
jayd16 2 hours ago [-]
Even if that was true, you'd have to still prove it has emerged.
krackers 54 minutes ago [-]
What would be your test to determine that?
slopinthebag 2 hours ago [-]
Why would "reasoning" be an emergent property of prediction?
js8 1 hours ago [-]
There's a lot of reasoning in the training data.
16 minutes ago [-]
mw888 1 hours ago [-]
Predict multiple outcomes, induct across them, refine.
hsn915 1 hours ago [-]
How do you predict without reasoning?
slopinthebag 44 minutes ago [-]
Where is the reasoning in linear regression?
danielbln 30 minutes ago [-]
Where is the reasoning in synaptic transmission?
bluegatty 2 hours ago [-]
"They’re foundationally incapable of always and consistently preventing prompt injection attacks. “Alignment work”, safety harnesses, and sandboxes all help to add barriers against the worst, but there are fundamental gap" ...
They seem to be very good at a lot of rudimentary best practices, more so than humans, but more accurately - if you run and audit pass with specific instructions ... they're very good at that.
I mean - it's what they're the best at which is applying 'fuzzy heuristics' in a mechanical way. If can describe issues concisely, the patterns, the styles, the rules then LLMs can very mechanistically and methodologically grind through them.
I don't even see how this is controversial - without getting into 'what their reasoning means' - we can all agree that their synthetic reasoning is pretty good at narrow scales, and they've been 'trained by compilers' and are extremely good at spotting common patterns.
If you back that up with a lot of tokens ... they excel.
Designing architecture, that's difficult, but hammering away at all the 'known-knows across a system' especially to identify things ... they're pretty good at that.
hirvi74 3 hours ago [-]
> In the past year, agent harnesses crossed the “can it be done” rubicon.
Brother, I'm still in "Can you get it right?"-mode. What am I doing wrong? (Rhetorical, but advice welcomed).
dosisking 15 minutes ago [-]
There are two 'camps' with respect to AI.
One camp already knows that Neural Nets don't work and are a dead end.
The other camp hasn't yet figured out that Neural Nets don't work, but are convinced that they do (or eventually will), because they think everything always improves over time in a linear fashion.
jaggederest 31 minutes ago [-]
I'd be happy to screenshare with you if you like, we can work on something trivial or open source. Half an hour should be more than enough to see whether you're doing anything obviously self-sabotaging.
al_borland 3 hours ago [-]
I’ve found some success is small projects, with limited scope, in a greenfield.
I’m terrified to attempt agentic anything in the repo my job actually cares about. I triggered it once by accident, when the agent was first rolled out and enabled by default… it broke everything. Now I just use ask mode, and even that is wrong half the time, and once it goes wrong it just keeps getting worse.
I saw a post from Dave Plumber who vibe coded up a new cross platform task manager. He said his spec document for the AI was 107 pages long. So maybe what I’m doing wrong is not giving the AI a literal novel of spec.
applfanboysbgon 1 hours ago [-]
> He said his spec document for the AI was 107 pages long.
This sounds like programming but with extra steps that make it take longer with less reliability.
0x696C6961 55 minutes ago [-]
Ikr, at that point the code itself is a better way of encoding the information.
mw888 1 hours ago [-]
You're appealing to ambiguity. All you've said is you have failed—how is anyone supposed to know what went wrong?
simonw 2 hours ago [-]
Tell it to use red/green TDD and start things off with an already configured test suite, maybe with a single test that asserts 1+1==2.
Make sure it know how to run the tests before it starts writing any additional code.
Then set it a clear goal.
slopinthebag 2 hours ago [-]
Basically all the examples of LLM's building impressive things have been because they have human written tests to base the implementation on. If you have an LLM write the tests the results are far less impressive or valuable.
bharatsuthar 1 hours ago [-]
Yes and LLMs are known to cheat on tests written by them.
slopinthebag 40 minutes ago [-]
It's not always cheating either. They aren't intelligent, so they don't actually understand the purpose of the tests or can build them to define the actual semantics of the problem space. It's literally just next-token prediction based on the codebase and prompt. Cheating implies that they have agency, and ironically agents don't.
MattGaiser 3 hours ago [-]
What is “it” specifically and what languages are you using?
Rendered at 07:30:36 GMT+0000 (Coordinated Universal Time) with Vercel.
> Making software debuggable, maintainable, layered, and composable – that’s still quite a trick
Not really. I have been working on a mobile app for months, and I stopped even glancing at the code about two months ago.
150k LOC, around half of that in tests, and the AI still has no problem maintaining the code on my behalf.
Debuggable? It can add extensive instrumentation in seconds.
None of this requires expertise, prompting, or mention of TDD. It's the default.
Frankly I do not believe the author tried developing a large codebase fully agentic and without reviewing the code. I believe many here look at the code produced, deem it substandard, and go hands on.
> They’re foundationally incapable of always and consistently preventing prompt injection attacks
From Anthropic's article about the Auto mode:
> We commissioned an evaluation from a third party, Trajectory Labs, who tested different models within the latest publicly available versions of Claude Code and Codex as of July 17th 2026.1 They tested 72 indirect prompt injection scenarios held out from Anthropic
> In this evaluation, none of the 720 attack attempts succeeded against Claude Fable 5, Opus 5, or Sonnet 5 running auto mode. On the other hand, 5.83% of the attacks succeeded against GPT-5.6 Sol running Codex's Auto-review mode. Notably, this is greater than the 0.09% average attack success rate against our latest models running in bypassPermissions mode without additional safeguards. The tests showed a 19.03% attack success rate against GPT-5.6 Sol when running in Full Access mode
I'm sure someone is going to reply with how they do not trust Antrophic's research, but lacking other data, prompt injection appears to be largely solved already.
The error types and codes, it will produce to spec.
If you type 'make me that thingy' - yes, it's probably not going to do what you want, but if you give it spec and guidance, it usually will.
The 'interface design' ... not very good though.
But once the insanity ends LLMs will be packaged as tools for developers to use to boost their productivity, and we'll consider them as we do IDE's and debuggers and stuff. But we have to get through this hype cycle first.
Semantics. Prediction is the training objective. The ability to reason can be, and very arguably is, an emergent property of that.
They seem to be very good at a lot of rudimentary best practices, more so than humans, but more accurately - if you run and audit pass with specific instructions ... they're very good at that.
I mean - it's what they're the best at which is applying 'fuzzy heuristics' in a mechanical way. If can describe issues concisely, the patterns, the styles, the rules then LLMs can very mechanistically and methodologically grind through them.
I don't even see how this is controversial - without getting into 'what their reasoning means' - we can all agree that their synthetic reasoning is pretty good at narrow scales, and they've been 'trained by compilers' and are extremely good at spotting common patterns.
If you back that up with a lot of tokens ... they excel.
Designing architecture, that's difficult, but hammering away at all the 'known-knows across a system' especially to identify things ... they're pretty good at that.
Brother, I'm still in "Can you get it right?"-mode. What am I doing wrong? (Rhetorical, but advice welcomed).
One camp already knows that Neural Nets don't work and are a dead end.
The other camp hasn't yet figured out that Neural Nets don't work, but are convinced that they do (or eventually will), because they think everything always improves over time in a linear fashion.
I’m terrified to attempt agentic anything in the repo my job actually cares about. I triggered it once by accident, when the agent was first rolled out and enabled by default… it broke everything. Now I just use ask mode, and even that is wrong half the time, and once it goes wrong it just keeps getting worse.
I saw a post from Dave Plumber who vibe coded up a new cross platform task manager. He said his spec document for the AI was 107 pages long. So maybe what I’m doing wrong is not giving the AI a literal novel of spec.
This sounds like programming but with extra steps that make it take longer with less reliability.
Make sure it know how to run the tests before it starts writing any additional code.
Then set it a clear goal.