AEvo: A Meta-Editing Framework for Agentic Evolution

A new paper on arXiv (2605.13821) introduces AEvo, a harnessed meta-editing framework that treats agentic evolution as an interactive environment. Agentic evolution iteratively generates candidates, evaluates them, and uses feedback to guide search. Existing methods are either fixed hand-designed procedures or general-purpose agents that can drift. AEvo uses a meta-agent that observes accumulated evolution context—including candidates, feedback, traces, and failures—as a process-level state, and acts by editing the mechanism driving future evolution rather than directly proposing candidates.