Anthropic says it taught AI agents how to 'dream'

Anthropic has introduced a groundbreaking technique called "dreaming" aimed at enhancing the capabilities of autonomous AI agents. This announcement was made during the company's developer conference on Wednesday, marking a significant step in their mission to revolutionize software engineering and knowledge-based work with increasingly independent tools. The "dreaming" process is designed to improve an AI system's memory by facilitating evaluations between operational sessions. By analyzing past behaviors and identifying patterns, the technique helps AI agents optimize their work methods and minimize errors. Currently, this innovative feature is available as a research preview, requiring developers to apply for access. In recent months, Anthropic has experienced a surge in revenue, largely driven by the adoption of its Claude Code service and related products, which assist software engineers in launching agents for long-term coding tasks. The startup is looking to extend its AI capabilities beyond software engineering into industries such as finance and law. Enhancing agents' ability to recall and learn from previous tasks could significantly improve their accuracy and productivity, thereby increasing their value for clients. Additionally, the announcement coincided with a thought-provoking essay by Anthropic co-founder and policy head Jack Clark, who suggested there is a 60% likelihood that advanced AI models will be capable of autonomously training their successors by the end of 2028. Clark referred to the current surge in research and the emergence of new products from leading AI labs. Although he did not specifically mention "dreaming," this technique aligns with the company's broader goal of developing AI models that can function independently. Clark noted, "This significant rise in the duration that AI systems can operate autonomously aligns perfectly with the rapid development of agentic coding tools," emphasizing the shift towards AI systems that perform tasks on behalf of individuals, working autonomously for extended periods. On the same day, Anthropic also transitioned two Managed Agents tools from research preview to public beta, enhancing user accessibility. One of these tools employs rubric-style outcomes to guide agents, while the other enables the delegation of tasks to multiple sub-agents.