PROTEUS Graph-Based Idea Evolution Through Autonomous AI Agents

This technical note presents PROTEUS, a system architecture for autonomous idea evolution in which concepts are treated as organisms within a graph-based ecology. Five specialized AI agents operate on a shared directed graph: PROMETHEUS (divergent ideation), METIS (refinement), MOMUS (adversarial critique), ANANKE (fitness-based selection and pruning), and HARMONIA (cross-branch synthesis and pattern detection). Ideas undergo mutation, crossover, selection, and fossilization through formally defined operators, with full provenance tracking enabling reconstruction of any concept’s evolutionary lineage. The architecture introduces graph-native evolution with typed cross-links (as opposed to tree-based structures), multi-parent synthesis via crossover operators, a fossil record serving as within-run memory, and ecological selection pressure through speciation with niche-specific budgets. A controlled experiment compares four conditions: human-only ideation, human–AI collaborative ideation, full autonomous multi-agent evolution, and a single-agent baseline. The experimental design encompasses 78 runs with blinded evaluation, anchored rubrics, inter-rater reliability measures, and a pre-registered statistical analysis plan. Additionally, a metapopulation architecture (KOSMOS/DAIMON) is proposed in which full-scale exploration instances spawn focused sub-instances through a biological budding mechanism, creating a two-level evolutionary dynamic grounded in metapopulation theory. Part of the DAEDALUS Cognitive AI Agent Architecture project at the University of Oradea. This document is a defensive public disclosure intended to establish prior art and timestamp the PROTEUS concept, architecture, and experimental design. It establishes priority for the following novel contributions: (a) a five-agent architecture mapping distinct evolutionary pressures (generation, refinement, critique, selection, synthesis) to specialized AI agents operating on a shared graph; (b) a formal evolutionary model with typed mutation and crossover operators, a five-dimensional fitness function, speciation with niche budgets, and a fossil record as within-run memory; (c) the KOSMOS/DAIMON metapopulation architecture implementing two-level evolution where agents handle micro-evolution within instances while humans control macro-evolution across instances through freeze-bud-compete-merge-reintroduce cycles; (d) provenance as an experimental instrument enabling full lineage reconstruction for post-hoc analysis of how ideas evolve; (e) a controlled experimental design with four conditions, pre-registered statistical analysis, blinded evaluation, and reproducible run-bundle format; and (f) the principle of resource allocation as selection pressure (investment strategy rather than rate limiting).