Anthropic has updated its agent management platform. The Claude service now features "dreaming" capabilities, performance evaluation, and orchestration of multiple bots.
Dreaming allows agents to analyze past sessions and identify patterns in their operations. The system detects recurring errors and successful scenarios, enabling it to autonomously refine algorithms. Users can either set up automatic memory updates for the AI or manually approve changes.
Developers have also introduced Outcomes and Webhooks tools. Now, application creators can establish quality criteria for agent performance. A separate "evaluator" checks the results against the rules independently of the main task execution process. If the outcome does not meet the standards, the agent works on correcting the errors.
According to Anthropic, Outcomes has improved the efficiency of complex task execution by 10%. The quality of document generation in .docx and .pptx formats has increased by an average of 9%.
The third update involves the orchestration of multiple agents. The "main" AI can delegate subtasks to specialized agents: one analyzes code deployment history, another reviews error logs, and a third handles support tickets. All of them work concurrently on a shared set of files.
Harvey and Netflix Test Updates
Partners of the company are already testing the new features. The Harvey platform uses Dreaming for preparing legal documents, resulting in a sixfold increase in task completion speed. The Netflix team has employed a multi-agent system to analyze application build logs.
The Dreaming feature is available in preview mode upon request. Outcomes and multi-agent orchestration have entered public beta testing.
Recall that on May 4, Anthropic co-founder Jack Clark predicted the emergence of "self-improving AI" by 2028.