Agent swarms and the new model economics

The big claim here is deliciously dramatic: a new team of AI helpers did far better than the old one on a tough coding challenge, getting to 80% on a hidden test in four hours while the previous version apparently melted down before hour two. The company says the trick is simple in theory: let the smartest AI do the planning, then hand the smaller jobs to cheaper, faster AIs. They also bragged about building a brand-new system to manage the flood of changes, because their old setup topped out at about 1,000 code changes per hour, while the new one allegedly screams along at 1,000 per second.

But the real show is in the comments, where readers split into two camps: future-believers and hype police. One excited fan called it a glimpse of the future, the kind of wild experiment people would have laughed at a couple of years ago. Others were much less starry-eyed. The sharpest side-eye came from people asking the obvious question: if the AI was trained on lots of public code, is rebuilding SQLite really proof of genius, or just a fancy open-book exam? Another popular complaint was basically, “Cool demo, where’s the actual useful product?” One commenter joked that these posts keep flexing the power of the process while admitting the resulting software still kind of stinks.

And then there’s the classic tech-drama subplot: show us the goods. Some readers said the most interesting part isn’t the flashy benchmark, but the behind-the-scenes machinery—and they’re annoyed that the company teased it without releasing the code. In other words: the swarm may be cooperating, but the comment section definitely isn’t.