Claude's new "Dreaming" feature is designed to let AI agents learn from their mistakes

Anthropic expanded Claude Managed Agents with new tools aimed at reliability and coordination.

Dreaming, now in research preview, reviews past sessions to find patterns and reusable lessons for agent memory.

Outcomes, in public beta, evaluates work against rubrics and can retry tasks when results miss the bar.

Multiagent orchestration lets a coordinator split work across specialized agents in parallel, also in public beta.

The update is meant to help developers build production agents that learn from experience, self-check outputs, and coordinate more effectively.