kadath

KADATH is a kernel for Agentic Darwinian Adaptation, Tooling, and Heredity. It takes a goal and uses repeated generations of agents that attempt the goal, are independently graded against a measurable benchmark, learn from the strongest performers, mutate, reproduce, and cull weak approaches. The system separates a kernel that owns the run, benchmark, deadlines, containers, evidence, grading formulas, population selection, database, Git lineage, recovery, exports, and cleanup from organisms that perform the goal.

Each organism genome contains a complete editable agent framework, including its system prompt, Python implementation, tools, dependency declarations, and supporting files. The evolutionary control plane remains outside the genome so an organism can improve itself without gaining the ability to rewrite its own scores, objective, scheduler, lineage, or isolation rules. The Architect converts the goal into a complete benchmark, and approval locks hashes of the objective, Architect output, tool manifest, and runtime configuration.

During an epoch, each checked genome is mounted read-only into its own parent container, while the organism receives a persistent writable workspace, ranked memory, artifact storage, a scoped model gateway token, approved optional capabilities, and a hard deadline. The Grader reviews a frozen evidence boundary rather than the agent's live workspace, and the kernel validates evidence references and applies locked formulas itself. After every graded non-final epoch, KADATH divides the population by verified outcome and rank, preserving top performers, allowing middle agents to mutate or remain unchanged, and culling failed or lowest-ranked successful agents.