Harnessing Agentic Evolution

Agentic evolution has emerged as a powerful paradigm for improving programs, workflows, and scientific solutions. It iteratively generates candidate artifacts, evaluates them against a task objective, and uses the resulting feedback to guide subsequent evolution. Existing methods typically instantiate this paradigm either through fixed hand-designed procedures that are modular but rigid, or through general-purpose agents that flexibly integrate feedback but can drift as context grows, leaving long-horizon search vulnerable to local optima. Both routes share a deeper limitation: long-horizon evolution accumulates candidates, feedback, traces, and failures over time, but lacks a stable interface for organizing this evidence and revising the mechanism that drives future evolution. We therefore formulate agentic evolution as an interactive environment, where the accumulated evolution context becomes process-level state. A meta-agent acts on this environment not by directly generating the next candidate, but by editing the mechanism that controls how future evolution proceeds. We introduce AEvo, a harnessed framework for meta-editing agentic evolution. It standardizes the evolution environment and provides a unified interface for observing accumulated evidence and editing the mechanism that drives future evolution. This lets AEvo revise both hand-designed procedures and agent operating contexts, reducing local-optimum risks in long-horizon evolution. Empirical evaluations on agentic and reasoning benchmarks show that AEvo outperforms 5 evolution baselines, achieving a 26% relative improvement over the strongest baseline. Furthermore, across 3 open-ended optimization tasks, AEvo outperforms 4 evolution baselines and achieves state-of-the-art performance under the same iteration budget.