KADATH Evolves Agent Populations to Solve Goals Through Competitive Selection 🔗
Each generation breeds improved autonomous agents by mutating prompts, code, and tools under fixed benchmarks
KADATH implements an evolutionary runtime that treats agent frameworks as genomes subject to selection pressure. Given a goal, it spawns a population of agents—each with editable system prompts, Python logic, tools, and dependencies—and grades them against a measurable benchmark.
Top performers inform the next generation through mutation and reproduction, while weaker agents are culled. Over epochs, the system converges on high-performing agent designs without relying on manual prompt engineering. In a demonstrated ten-epoch run, the leading cohort’s best score jumped from 18 to 91, with the top-five median rising from 8 to 77, indicating broad population improvement rather than individual outlier success. The kernel preserves evaluation integrity by isolating fitness scoring, scheduling, and lineage from the evolvable agent genome, preventing self-modification of objectives. Organisms remain read-only during epochs, only adapting in post-grade mutation phases. Built in Python and leveraging smolagents for core agentics, KADATH provides a reproducible, containerized pipeline with Git-based lineage tracking and recovery. It shifts focus from optimizing single agents to cultivating adaptive populations that collectively solve complex, open-ended goals through competition.
The catch: As an early-stage project with no open issues and minimal external adoption, its real-world efficacy beyond toy benchmarks remains unproven at scale or in noisy, dynamic environments.
Why this leads today i3T4AN/KADATH provides a new evolutionary runtime enabling autonomous agent swarms to self-improve across reproducible epochs, offering developers a foundational tool for optimizing agent systems in practice.
- Researchers evolving LLM agents for scientific hypothesis generation
- Teams automating prompt and toolchain optimization for repetitive tasks
- Developers benchmarking agent robustness against shifting success criteria
Source: i3T4AN/KADATH — based on the project README.