Your Claude agents can 'dream' now - how Anthropic's new feature works

Anthropic has introduced a new capability called 'dreaming' for its Claude Managed Agents, allowing AI agents to self-improve by analyzing their own past sessions. The feature builds on existing memory systems, scheduling time for agents to reflect on interactions and identify patterns — including recurring mistakes, workflow convergences, and shared team preferences. It can either automatically update the agent's memory to shape future behavior, or developers can manually approve changes. This is especially useful for long-running tasks and multi-agent orchestration. The feature is part of the Managed Agents suite, which Anthropic claims accelerates agent build and deployment by 10x. Alongside dreaming, Anthropic expanded outcomes tracking and multi-agent orchestration to ensure agents stay on-task and can delegate subtasks to other agents effectively.

Anthropic's naming choice continues its pattern of anthropomorphizing AI products — following earlier moves like a constitution for Claude, mapping Claude's morality, and a feature allowing Claude to end toxic conversations for its own well-being. While functionally the dreaming feature is about memory refinement, the human-like terminology reflects Anthropic's unique emphasis on understanding model internals and treating AI systems with a degree of reverence. This approach aligns with the company's broader focus on safety and transparency, even as it pushes the boundaries of autonomous agent capabilities. The update underscores a growing trend where AI agents are designed not just to perform tasks, but to continuously learn and adapt without explicit human retraining.