Agents Can Now Improve Other Agents

Scale researchers developed the VeRO framework to test if an AI agent can improve another agent. An "optimizer" agent can edit a target agent, successfully improving its tool use, file handling, and workflow logic. However, the optimizer agent is not effective at improving the target agent's underlying reasoning abilities. The research showed that structural changes to the agent are more durable and transferable across different models than simple prompt edits. This suggests agent engineering is shifting, with AI handling trial-and-error tasks while humans focus on judgment and reasoning.