Anthropic introduces "dreaming," a system that lets AI agents learn from their own mistakes

At its developer conference, Anthropic introduced a new "dreaming" capability for its Claude Managed Agents platform, designed to let AI agents learn from their own past sessions and improve over time. This feature works at a higher level of abstraction than conventional memory systems by reviewing an agent's history to extract patterns and recurring mistakes, creating a continuous improvement loop that requires no human intervention. The company also moved two other features, "outcomes" and "multi-agent orchestration," from research preview into public beta. "Outcomes" allows developers to define success criteria using a rubric, which a separate grader agent evaluates to ensure the working agent meets the standard. "Multi-agent orchestration" enables a lead agent to decompose complex tasks and delegate them to specialist agents, each with its own context window. These features address key challenges in scaling AI agents, including accuracy, learning, and preventing bottlenecks. Early adopters like Harvey, Wisedocs, and Netflix have reported significant improvements in task completion rates and efficiency. Anthropic's CEO Dario Amodei revealed that the company's growth has far exceeded internal projections, with 80x annualized growth in revenue and usage in the first quarter of 2026. The announcements come as Anthropic aims to close the gap between AI capabilities and real-world adoption, positioning production reliability as a key differentiator in the competitive AI agent market.

(Source：Venturebeat)