Anthropic just gave its Claude Managed Agents a new trick: they can "dream" while idle. In a research preview, agents periodically review past sessions during downtime, spotting recurring errors and team preferences, then automatically organize those insights into a memory store. Users can choose to let the bot write directly to the memory or require a human review.
Two other features also entered public beta: Outcomes lets you define scoring criteria, with an independent judge model evaluating the agent's work in a separate context — if it fails, the agent redoes the task. Internal testing showed up to a 10-percentage-point improvement in task success rates. Multi-agent orchestration lets a lead agent break tasks into subtasks and hand them off to child agents, each with its own model and tools, all sharing a file system.
Legal AI company Harvey tested the dreaming feature and saw the completion rate for complex legal documents jump about sixfold in trials.