Claude Managed Agents Now Dream Between Sessions - Code with Claude 2026 Announcements

Anthropic has introduced a groundbreaking feature called 'dreaming' for its Claude Managed Agents during the Code with Claude event. This automated background process allows agents to analyse their completed sessions overnight, identifying patterns, recurring mistakes, and successful workflows to enhance future performance. The system essentially enables agents to curate their own memory and return each day with improved capabilities, requiring no manual intervention from users. The dreaming functionality works by mining agent sessions for valuable insights, including workflows the agent has converged on and preferences shared across teams. This marks a significant shift towards autonomous agent improvement, where the AI system actively learns from its own experiences during downtime periods. Alongside dreaming, three additional capabilities have moved to public beta. The 'Outcomes' feature allows users to establish success criteria, with a separate Claude instance grading outputs until they meet the required standards. Anthropic's internal benchmarks demonstrate up to ten percentage points improvement on challenging tasks using this approach. The update also introduces multiagent orchestration, enabling lead agents to delegate specialised subtasks to other agents, and webhook support for event-driven triggers rather than continuous polling. These enhancements collectively represent a substantial advancement in agent autonomy and effectiveness, with dreaming being particularly noteworthy as the first system designed to compound agent quality without active oversight.