i3T4AN/KADATH
An open-source runtime that evolves AI agents across generations by having them compete for a goal
KADATH takes a user-defined goal and spins up a population of AI agents that are graded against the same locked benchmark each round, keeping the strong performers and replacing the weak ones to breed progressively better agents. What evolves is not just a prompt but the agent's entire framework, including its Python code, tools, and dependencies, while a separate kernel handles execution, grading, and selection without evolving itself. In a published ten-epoch run, the top-five median fitness rose from 8 to 77 and the best score rose from 18 to 91.
What it does
- The user provides a goal, epoch duration, population size, and number of epochs, then a dedicated Architect model designs a scoring benchmark that must be approved before the run starts
- Each agent runs inside its own isolated Docker container with its code mounted read-only, logs its activity, and must produce a result within a fixed deadline
- After each epoch, a Grader scores agents from frozen evidence; the top 30% are preserved unchanged, the middle group decides for itself whether to mutate, and the lowest performers are culled and replaced
- A Tweaker analyzes what made elite agents effective and sets a reproduction brief, which a Birther uses to create new child agents (genomes) that fill out the next generation's population
- Across a published ten-epoch run, the top-five agents' best score rose from 18 to 91, median rose from 8 to 77, and lowest score rose from 1 to 71
Why it matters
It offers a concrete way to improve AI agent quality through repeated competition and selection rather than betting everything on a single prompt or agent design. Its detailed grading, isolation, lineage, and recovery machinery makes it a useful reference for anyone wanting a reproducible framework for agent experimentation.
Terms in this repo
- genome · the complete evolvable bundle of an agent's system prompt, code, tools, and dependencies
- epoch · one generation cycle in which the agent population attempts the goal and gets graded
- Architect/Grader/Tweaker/Birther · specialist model roles responsible for designing the benchmark, scoring, analyzing elites, and producing new agents
- fitness · an agent's performance score under the locked benchmark
Repository description (English)
Evolutionary multi-agent runtime that breeds, evaluates, and improves autonomous agents across reproducible epochs to converge on optimization of a goal.
Open on GitHubTrending repos
- vorssaint/vorssaint-utilsOne free menu bar app replaces a dozen paid Mac utilities
- Alishahryar1/free-claude-codeA local proxy that lets coding AI agents run on 49 free or cheap model providers instead of one paid service
- freestylefly/awesome-gpt-image-2A library of 532 reverse-engineered prompts that turn GPT-Image2 into a predictable image-making tool
- block/buzzAn open-source workspace where humans and AI agents chat, code, and review in the same rooms
- NousResearch/hermes-agentNous Research's Hermes is an AI agent that gets smarter the more you use it
- virgiliojr94/book-to-skillA tool that turns technical book PDFs into on-demand reference skills for AI coding agents
- VoltAgent/awesome-agent-skillsA single hub collecting over 1000 'how-to' manuals that make AI coding assistants act like experts
- anthropics/claude-plugins-communityA shared shelf where anyone's Claude add-ons get listed for install