KADATH: a day-old repo lets agent populations breed, mutate, and get culled across epochs

A GitHub repo called KADATH dropped yesterday and it’s doing something most agent frameworks skip: it treats populations of AI agents like organisms that evolve over time.

Here’s the mechanism. You spin up a population of agents built on Hugging Face’s smolagents library. KADATH scores each agent against whatever objective you set, then keeps the top 30 percent. The middle tier gets to mutate, meaning they tweak their own instructions or approach. The bottom tier gets culled. Rinse and repeat across epochs. It’s selection pressure applied to software, and the repo is bare bones right now, sitting at 105 stars.

Why it matters: most companies running agents treat them as static. You write a prompt, maybe you tune it once, and it runs the same way forever. KADATH suggests a different model where agents improve themselves over repeated cycles without you rewriting code every time performance drifts. If you’re dispatching dozens of agents to handle customer queries or data tasks, you want the system to figure out which approaches work and which don’t, then double down on what works. That’s conceptually close to how we think about dispatch and scoring inside the Omni Command Centre, where agent performance feeds back into routing decisions.

The caveat: this is a day-old repo with no production track record. It’s an experiment, not a product. But the idea of culling weak agents and letting strong ones mutate is worth watching, especially if you’re already running multi-agent workflows and wondering why some agents drift or fail quietly while others hold up. It won’t replace your current setup tomorrow, but if this pattern proves stable, it could shift how we architect agent systems that need to stay sharp over months, not just days.