Anthropic has added a new feature called "dreaming" to its AI agent infrastructure, Claude Managed Agents. The feature lets agents review their own sessions, extract patterns, and improve their behavior over time — all while using less memory.
https://twitter.com/claudeai/status/2052067399088664981
Dreaming analyzes an agent's past sessions to find recurring patterns and refines its internal instructions accordingly. The result is an agent that can handle complex tasks with minimal prompting and a smaller memory footprint. Anthropic says this is a step toward AI systems that continuously self-optimize without human intervention.
The feature arrives as Anthropic recently secured a deal with SpaceX for access to the Colossus supercomputer, dramatically boosting the compute available for Claude agents. The move comes amid explosive growth for Claude — monthly active users jumped 658% year over year.
Users can decide how much control to keep: the system can update memory automatically, or changes can be reviewed by a human before they take effect. Dreaming is available now as a research preview in Claude Managed Agents. Developers can apply for access through Anthropic's website.