Anthropic teaches AI agents how to ‘dream’

At the Code with Claude developer conference on May 6, Anthropic demonstrated a new feature called Dreaming for Claude Managed Agents (CMA). CMA is a cloud-based AI agent platform that provides off-the-shelf infrastructure to help developers quickly build and deploy automated AI systems. This is considered a solution that shortens the time to develop AI agents 10 times more than usual, by handling complex execution environments and infrastructure.

Dreaming works through the process of reviewing a list of recent events, identifying specific issues, selecting them and storing them in memory for future tasks and interactions. According to Arstechnicathis “daydreaming” process is considered very important, in that it overcomes the drawback of the limited context window present on large language models (LLM), especially important information that can be lost in long-term projects.

On the chatbot chat side, many models use a process called compression. Accordingly, long conversations are periodically analyzed and an attempt is made to remove irrelevant information from the context window while retaining what is truly important for the conversation, project, or task at hand. With Dreaming, scanning is repeated periodically, where past sessions and memory stores can be analyzed across multiple agents. Users can choose between automatic processing or live review of changes in memory.

“Dreaming helps improve the limitations of AI models that use single agents, such as repetitive errors, the workflow of multiple agents at the same time, and the ability to share data. It also helps restructure memory, maintaining transparency,” Anthropic said.

Dreaming is currently in beta research, limited to developers who have pre-registered for CMA on the Claude platform. According to Anthropic, CMA is a higher-level alternative to building directly on top of the Messages API, described as a “pre-built, configurable agent system running on managed infrastructure,” designed for users who want multiple agents to work on a task or project towards some endpoint in minutes or hours.

Anthropic’s revenue has soared in recent months as software engineers embraced its Claude Code service and programming agent products, despite being at odds with the Pentagon. The company’s value on the secondary market reached $1,000 billion, surpassing the $880 billion figure of rival OpenAI.

Theo Business InsiderAnthropic’s “daydreaming” feature also sheds light on a notable essay posted to the blog earlier this week by Jack Clark, co-founder and head of policy at Anthropic. In it, without mentioning Dreaming, he hypothesizes a 60% chance that advanced AI models will automatically train their own next generation by the end of 2028.

“The dramatic increase in the time it takes AI systems to operate independently correlates closely with the explosion of agent programming tools,” Clark wrote. “This can be seen as the ‘self-commercialization’ of AI systems in performing work on behalf of humans.”