Anthropic has introduced a new system called Dreaming for its Claude Managed Agents platform. This technology allows AI agents to independently analyze their experience, learn from past mistakes, and improve efficiency in future tasks. Unveiled at the Code with Claude conference in San Francisco, this solution aims to develop the agents' self-correction capabilities without retraining the model. This is reported by Ixbt.com news.
Unlike simple memory, the Dreaming system operates as a separate process in a "sleep mode." The system analyzes past sessions, identifies recurring errors and successful solutions, and converts them into structured guidelines for future work. Anthropic researcher Alex Albert noted that this process is similar to human professional development, where the agent creates custom "playbooks" for itself.
The company also introduced Outcomes and Multi-Agent Orchestration technologies. Outcomes allow agents to automatically verify their work based on defined quality criteria, while Multi-Agent Orchestration helps distribute complex tasks among several specialized agents. This approach serves to prevent quality degradation in long-running sessions.
Anthropic CEO Dario Amodei stated that the company's growth has significantly exceeded expectations, with annual revenue increasing 80-fold. To meet the demand for computing power, the company is expanding its infrastructure through the Colossus data center in partnership with SpaceX.
Read “Zamin” on Telegram!
