Claude Managed Agents add dreaming, orchestration

Anthropic is turning Managed Agents into a real control plane for production agents, not just a wrapper around tool calls.

• –Dreaming is the interesting piece: it converts accumulated session history into curated memory, which is where long-lived agents usually start to decay.
• –Outcomes is a cleaner way to do eval-driven automation because the rubric lives outside the agent’s own context, reducing self-justifying loops.
• –Multiagent orchestration matters more than the marketing language suggests; it gives teams a native lead/subagent model instead of forcing them to build queues and fan-out logic themselves.
• –Webhooks are the practical glue for production use, since completion events beat polling when agents run for minutes or hours.
• –The feature set is still early-stage, so the real question is whether Anthropic can keep the abstraction sharp as teams push it into brittle, high-stakes workflows.