AI Pulse · AI Trends Pulse

A new project called KADATH dropped on GitHub yesterday and it’s worth watching if you run multi-agent workflows. It’s an experimental runtime that treats AI agents like a breeding population. The system scores them, keeps the ones that perform, and culls the rest across what it calls “epochs.” Think natural selection, but for code assistants.

The repo sits at 105 stars after one day, which means it’s very early. It’s built on Hugging Face’s smolagents framework, so it’s not a closed black box. What makes it interesting is the structural idea: instead of hand-tuning one agent or running a fixed set in parallel, you let a population evolve based on real task performance. The approach shares DNA with dispatch-gate scoring systems that route work to the agent most likely to succeed, but KADATH pushes further by actually removing underperformers and spawning new variants.

This sits alongside the skill-pack marketplaces cropping up in the Claude Code plugin world. Those let you buy or share pre-built agent skills. KADATH goes a different direction: it generates and tests agents on the fly. Both ideas point to the same pressure, companies want agents that adapt without constant human re-tuning.

What it means for operators

If you’re running agents in production, the question isn’t whether KADATH itself is ready (it’s not). The question is whether you have any way to measure which agents or prompts actually work in your environment and which waste time. Most teams still guess. A command centre that logs task outcomes, scores agent performance, and routes work accordingly is how you avoid that guessing game. That’s the kind of infrastructure we build into the Omni Command Centre, where scoring and dispatch happen automatically, not as a side project.

You can follow the KADATH repo if you want to see where evolutionary agent design heads next. Just don’t confuse experimental with deployable yet.