MetaEvoForge: Learning to Improve Agents Faster Across Task Environments

AI systems that improve other agents must learn how to make effective changes and carry that ability into new task environments. We study whether a shared metaagent can develop reusable improvement strategies from its experience of modifying taskagents. MetaEvoForge couples package-level agent evolution with the self-evolution of an independent improver, while keeping language-model weights fixed. It records parent behavior, a scoped package edit, and child behavior on matched tasks, then consolidates this evidence across environments to update the metaagent's persistent improvement policy. We compare trained and untrained improvers by evolving a common root on held-out task pools under a fixed search budget. The results show task-dependent improvements in final quality and instances of earlier discovery of high-scoring candidates. Behavioral analysis links these outcomes to reusable repair principles and the specialization of editing strategies, while revealing limits to their generalization. This work provides an empirical basis for studying how accumulated intervention experience can improve the process of improving agents across task environments.