Anthropic has introduced a new feature called 'dreaming' for its Claude Managed Agents, enabling AI agents to 'self-improve' by reviewing and learning from past sessions. This update builds upon existing memory capabilities, marking a significant step in autonomous AI learning.
Launched on April 8, Claude Managed Agents provide a suite of APIs designed to streamline the complex process of building and deploying AI agents, allowing teams to launch agents at scale up to ten times faster. The new 'dreaming' feature empowers agents to analyze past interactions, identify patterns, correct mistakes, and optimize workflows.
Specifically, the 'dreaming' feature schedules dedicated time for agents to reflect on and learn from their past interactions. Once activated, it can either automatically update an agent's memories to shape future behavior, or users can select which proposed changes to approve. According to Anthropic, 'dreaming' uncovers patterns that a single agent might miss, including recurring errors, convergent workflows, and shared team preferences. It also restructures memory to maintain a high signal-to-noise ratio as it evolves, which is particularly beneficial for long-running tasks and multi-agent orchestration.
During the 'Code with Claude' event, Anthropic product team members demonstrated the feature, referring to completed runs as 'finished dreams.' Concurrently, Anthropic also expanded two existing features: 'outcomes,' which keeps agents on task, and 'multi-agent orchestration,' which manages task delegation among agents. The company stated that these updates are designed to ensure agent accuracy and continuous learning.
Notably, Anthropic's choice of the anthropomorphic term 'dreaming' aligns with its historical approach to naming its products. Functionally, 'dreaming' refines an agent's reference pool, ideally making it more proficient at assigned tasks. Anthropic has consistently anthropomorphized its models and products; for instance, in January, it published a 'constitution' for Claude, intended to guide the chatbot's decision-making and define its ideal 'entity' status. Language within that document even suggested Anthropic was preparing for Claude to potentially develop consciousness.
