Beyond Self-Evolution: Meta-Evolving Agents for Cross-Task Agent Improvement

Agents are increasingly improved by other agents. We study meta-evolution as a way to make this process more effective across task environments: an improver learns how to make effective changes from its accumulated experience of modifying agents. Our central question is whether this experience yields reusable improvement strategies that transfer to unseen tasks. We propose MetaEvoForge, which couples task-level evolution of TaskAgents with the self-evolution of an independent improver, MetaAgent, while keeping all language-model weights fixed. Using asynchronous tree search across 62 training task environments, MetaEvoForge collects 768 improvement pairs per training run. Each pair records a scoped package edit and the parent and child agents' behavior on matched tasks. The shared MetaAgent consolidates this evidence across environments to update its persistent improvement policy. We evaluate evolved and unevolved improvers on held-out task suites, starting from the same root agent under a fixed search budget. Evolved improvers yield task-dependent gains in final quality and discover high-scoring candidates with fewer tokens. These findings show that improvement experience accumulated at scale can itself transfer across task environments.