I haven't seen an ounce of evidence that a LLM can replace a deeply experienced engineer. The harness turns coding into sculpting, which happens to be how I coded pre-LLM, but much slower, so I feel very much empowered by them. But I'm still definitely guiding the entire project. I haven't once felt like it wasn't required for me to be in the loop.
The people and companies leaning into fully autonomous agents are high on their own supply, in my opinion. They're kicking back in their beach chairs as their pipelines spit out Stüssy S after Stüssy S, with massive architectural flaws and attack surfaces, just lighting gobs of money on fire.
Yesterday, I created a fully functional POC for something really cool, it took me all day as I reshaped the agent's rough boilerplate ideas into usable components, and I never once hit a session limit on my $100 month Claude sub. I spent the majority of my time thinking about how I needed to prompt to turn what was in my head into working and secure code. You can't just give the agent a vague idea and expect anything less than a dumpster app.
It's probably enough to fool C Suite people into believing the AI apocalypse is coming, which is the crux of the problem, and what is fueling what is certainly a gigantic bubble — but when it comes down to it, shitty software is shitty software. To fix it, you have to know what good software looks and feels like under the hood, and why it's shitty or feels bad to use. It might start up, but mutability, staying up, remaining stable, and remaining secure are very different stories.
The clock is ticking on everything that has been developed by a LLM with a novice user behind the prompts.