Anthropic has announced a new feature for Claude Managed Agents called “dreaming.” The feature is currently in a research preview and is designed to enable AI agents to independently improve their performance over time.
According to Anthropic, dreaming analyzes previous sessions and stored memory to recognize patterns, correct errors, and refine workflows.
Dreaming expands upon the existing memory system within Claude Managed Agents. While memory is designed to retain information during a task or session, dreaming focuses on processing and improving that memory between sessions. The system reviews past tasks, saved context, and interactions to determine which information remains relevant and which processes work well or poorly.
Limitations of context windows
According to SiliconANGLE, Anthropic positions Dreaming as a solution to the limited context windows of large language models. In long-term tasks, context is lost when previous interactions no longer fit within the active working memory. Many AI systems therefore use techniques that periodically summarize conversations, retaining only relevant information. According to Anthropic, Dreaming goes further by not only analyzing individual conversations but also recognizing patterns across multiple sessions and agents.
The feature analyzes sessions and memory stores to identify recurring patterns, such as common mistakes or working methods that consistently yield better results. Additionally, dreaming can recognize shared preferences when multiple agents collaborate within the same environment.
An important part of dreaming is the restructuring of stored memory. According to Anthropic, this is intended to prevent memory systems from becoming cluttered with irrelevant or contradictory information over time. By actively reorganizing memory, the system aims to maintain the quality of stored context.
Developers can determine for themselves how often they dream. Additionally, it is possible to automatically apply memory updates or review changes manually first.
Dreaming is primarily intended for long-term workflows in which agents perform tasks over extended periods or collaborate with other agents. In these scenarios, the feature is designed to carry over experiences from previous sessions into new tasks without requiring constant manual intervention.
Agents distribute tasks among themselves
Anthropic explicitly links Dreaming to multi-agent orchestration, another new feature within Managed Agents. This allows a lead agent to distribute work among specialized sub-agents with their own models, prompts, and tools. The sub-agents can work in parallel on a shared file system, with their results consolidated within the context of the lead agent.
According to Anthropic, Claude Managed Agents is intended as an alternative for developers who do not want to build AI agents directly using the Messages API. For now, dreaming is only available through a developer access program. Outcomes, memory, and multi-agent orchestration are available as a public beta within Claude Managed Agents.