AI Can Now Hire Other AIs to Do Its Work — And It's Getting Weird

A research engineer named Florian Brand has published a piece arguing that the newest generation of AI models — he points to a model called GPT-6, nicknamed Astra — behaves differently from anything before it. Instead of answering one question at a time, it can break a big task into smaller ones and hand those off to helper AIs. Those helpers can then spawn their own helpers. Brand calls this "swarm" behavior, and he says that when it's set up correctly, it's impressive to watch.

The catch is that the model is still rough. It makes mistakes, and Brand describes it as "close to an alien mind" — meaning its reasoning can feel strange and hard to follow, even to experts. The tools wrapped around it (the software that tells it what to do) are holding it back. He compares this moment to early 2024, when the first reasoning models appeared and were clumsy too. The difference now is that the clumsiness is happening in systems that can act on their own.

There's also a money problem. Exploring swarm behavior properly costs between $10,000 and $100,000 in computing per attempt, so only well-funded labs can do it. That means the rest of us — including safety researchers, regulators, and ordinary users — are learning about these systems secondhand, from blog posts rather than independent testing.

Why should you care? If this trend continues, the AI you use at work may soon be a manager of dozens of invisible helpers rather than a single chatbot. That could mean far more gets done in far less time. It could also mean that when something goes wrong, nobody — not even the people who built it — can fully explain which AI made the bad call. Brand's closing message is blunt: if you have the budget and the skills to test these swarms, please do.