i3T4AN/KADATH

Evolutionary multi-agent runtime that breeds, evaluates, and improves autonomous agents across reproducible epochs to converge on optimization of a goal.

What it solves

KADATH is an evolutionary framework for creating high-performance AI agents. Instead of manually tuning a single prompt or agent configuration, it uses a Darwinian process of competition, selection, and mutation over multiple generations to evolve agents that are progressively better at achieving a specific, measurable goal.

How it works

KADATH operates through a "kernel" (the control plane) and "organisms" (the agents being evolved). The process follows a cyclical evolutionary loop:

  1. Benchmark Creation: An "Architect" model converts a user-defined goal into a strict, machine-readable benchmark with rubrics and anti-fraud rules.
  2. Population Execution: A population of agents (organisms) attempts the goal. Each agent has its own genome consisting of a system prompt, Python implementation, tools, and dependencies.
  3. Grading: A "Grader" model reviews frozen evidence (logs, artifacts, browser history) to extract facts and apply the benchmark's formulas to calculate a fitness score.
  4. Selection: The top 30% of agents are preserved. The middle cohort can reflect on their own performance and mutate. The lowest-performing agents are culled.
  5. Reproduction: A "Birther" creates new agents based on the elite performers, applying mutations suggested by a "Tweaker" model that analyzes what made the elites successful.
  6. Heredity: Children inherit their parent's genome, persistent workspace, and a filtered view of ancestral and population-wide memories.

Who it’s for

It is designed for developers and researchers who need highly optimized agents for complex tasks where the best prompt or framework configuration is not immediately obvious and can be discovered through automated iteration.

Highlights

  • Evolvable Frameworks: Evolves not just prompts, but the entire agent framework, including Python code and dependencies.
  • Strict Isolation: Agents run in secure, read-only containers with restricted resource limits and the kernel managing all credentials.
  • Evidence-Based Grading: Grading is performed on frozen, immutable evidence boundaries to prevent agents from cheating or manipulating their own scores.
  • Knowledge Heredity: Implements a sophisticated memory system where agents inherit ancestral records and can rate the utility of population-wide memories.
  • Automated Benchmarking: Uses a specialized Architect model to turn vague goals into verifiable, machine-readable metrics.

Related

  • Project
    alibaba/skill-up

    An evaluation and evolution tool for Agent Skills that allows developers to measure skill quality via declarative YAML cases and automatically improve them through a conversational repair loop.

  • Project
    A-EVO-Lab/a-evolve

    A universal infrastructure for autonomously evolving AI agents by mutating their workspace files (prompts, skills, and memory) to improve performance across any domain.

  • Project
    Optima-CityU/LLM4AD_Next

    LLM4AD Next is a Python toolkit that turns a natural‑language problem description into a full, runnable evolutionary‑algorithm search project. An interactive LLM consultant generates the code skeleton, configuration, and evaluator, then built‑in evolutionary methods (IslandGA, Diverse Island GA, EoH, etc.) automatically evolve the generated algorithm. It also offers an advisor to suggest evolution‑ready code blocks, long‑term memory support, Docker images, and an online demo.

  • Project
    alchaincyf/darwin-skill

    An iterative optimization framework for Agent Skills that uses a 9-dimensional rubric and a ratchet mechanism to ensure skills only evolve toward higher quality.