KADATH
An evolutionary runtime that breeds and grades populations of autonomous coding agents against a benchmark over repeated generations, evolving each agent's system prompt, code, tools and dependencies rather than just its output.
KADATH takes a goal and runs many generations of a population of agents at it. Each generation is graded against a measurable benchmark, the strongest performers are used to produce the next generation through mutation and reproduction, and weaker approaches are discarded, so the system searches for a good agent rather than betting everything on one prompt.
The system separates a kernel, which owns the run, benchmark, deadlines, containers, grading, database, lineage and cleanup and does not itself evolve, from the organisms, which perform the goal and whose complete framework, including system prompt, Python implementation, tools and dependencies, is what gets evolved; an organism's own repository is read-only while its epoch runs and can change only during the post-grade mutation phase. It is driven from an interactive terminal interface and needs an OpenAI API key to run.