Anthropic added three capabilities to Claude Managed Agents—"dreaming" (a research preview for scheduled memory review/self-improvement), "outcomes" (rubric/goal evaluation via a separate grader), and multiagent orchestration for delegating sub-tasks to specialists.