Anthropic’s Claude Managed Agents can now “dream,” sort of

SAN FRANCISCO—At its Code with Claude Developers conference, Anthropic announced what it calls the “dream” of Claude Managed Agents. Dreaming, in this case, is a way of going beyond the current experience and realizing the real things that need to be stored in “memory” to inform future activities and interactions.

Dreaming is the featured part of the survey and features Managed Agents on the Claude Platform. Managed Agents are another advanced way to build directly on the Messages API that Anthropic describes as “pre-built, flexible workhorses that are managed by the architecture.” They are designed for when you want multiple projects to work or a project to finish in a few minutes or hours.

Anthropic describes dreaming as a preprogrammed process, in which parts and memory stores are reviewed, and specific memories are stored. This is important because product windows are limited for LLMs, and important information can be lost during long projects. On the chat side, many models use a method called compaction, in which long conversations are analyzed from time to time, and models try to remove unnecessary information from the display screen and keep what is important for the conversation, project, or work.