GitHub radar
Kadath is an evolutionary runtime for creating superior AI agents
KADATH iteratively evolves a population of AI agents towards a solution to your task.
KADATH is an open source evolutionary multi-agent runtime written by single developer i3t4an on top of Hugging Face smolagents. It uses a biologically inspired approach to evolve solutions to tasks, where a population of agents compete over several epochs to solve your task, each epoch being evaluated independently with respect to a frozen benchmark, and the kernel then applying Darwinian principles of natural selection (preserving the top 30%, allowing the mid-tier agents to mutate according to the successful mutations of the top agents, and eliminating the bottom tier). Over a reported 10 epoch run, the maximum fitness increased from 18 to 91, demonstrating that the system improved over time through competition between agents, without the need to tune individual agents.
Why a vibe-coder should care
Rather than trying to tweak an individual agent’s prompts to find something that works, KADATH will automatically explore the solution space to find an optimal solution. Simply specify your task, the number of rounds and agents to use, and watch the system evolve to a better solution. You will need Docker and an OpenAI API key, but if you know how to run a shell script, you should be good to go!
▌ More finds