New in Claude Managed Agents: dreaming, outcomes, and multiagent orchestration (claude.com)

Claude Managed Agents has introduced "dreaming," a feature that enhances agent memory by reviewing past sessions to identify patterns and facilitate self-improvement. This capability allows agents to refine their memory and learning processes, either automatically or through user oversight, making them better suited for long-term projects and multiagent tasks. Alongside dreaming, the update also includes the introduction of "outcomes," which allows developers to set success criteria that agents independently use to evaluate and refine their outputs. This leads to significant improvements in task accuracy, especially in detail-oriented or subjective tasks, with reported success increases of up to 10 points in testing. The extension of multiagent orchestration enables a lead agent to delegate tasks to specialized subagents, enhancing efficiency and collaboration on complex workloads. This structure allows for real-time coordination and results in greater visibility into task execution. Companies like Harvey, Netflix, and Spiral by Every are already leveraging these advancements to boost task completion rates and maintain high-quality outputs. As these new features move into public beta, they promise to transform how teams use AI to deliver more effective and streamlined solutions.