Improvement Is Not Progress: Guarded Coevolution, Path Dependence, and Contagious Memory in AI Agents

Abstract

Language-model agents can adapt through prompts, memories, skills, tools, routers, evaluators, and environments without changing model weights. This paper reconciles recent work on adaptive environments, harness continual learning, and the fragility of memory-based self-improvement. It argues that coevolution increases adaptive capacity but does not imply monotonic progress, and proposes a Guarded Agentic Coevolution framework separating candidate variation, local adaptation, contextual stability, governed transfer, and authorized reliance. The framework adds multi-run and task-order stress tests, typed memory applicability, contagion monitoring, quarantine, non-compensable preservation, receiver-specific transfer, and rollback. Applications to JustitIA and IntegridAI are structural proposals under eval-first status, not evidence of empirical product improvement.