Claude has introduced several new capabilities for its Managed Agents platform, including dreaming, outcomes, and multiagent orchestration. Dreaming, now available as a research preview, allows agents to review past sessions, identify patterns, and refine their memory automatically or with user oversight. This process helps agents self-improve over time by curating and updating memory stores, making them more effective for long-running or collaborative tasks.

The outcomes feature enables developers to define success criteria for tasks, which agents then use to evaluate and refine their work. A separate grader assesses the results independently, allowing agents to self-correct without human intervention. Internal testing showed measurable improvements in task success and file generation quality. Developers can also integrate webhooks to receive notifications when outcomes are completed.

Multiagent orchestration allows a lead agent to divide complex jobs among specialized subagents that work in parallel and share context. This structure provides transparency through the Claude Console, where users can trace each agent’s actions. Dreaming is available in research preview, while outcomes, orchestration, and memory are in public beta.