Claude新功能:Dreaming、多代理编排与OutcomesNew in Claude Managed Agents: Dreaming, Multiagent Orchestration, and Outcomes
Claude Blog:Claude Managed Agents新增功能:Dreaming、多代理编排与Outcomes。
Anthropic为Claude Managed Agents推出研究预览版“Dreaming”。这个定时进程会回顾过去会话、提取模式并优化记忆,让代理能够随时间自我改进。与持久记忆结合,形成了强大的自我改进系统。
Outcomes允许开发者定义成功标准,由独立评估器判断结果,让代理不断迭代直到达标。该功能将任务成功率提升高达10个百分点,在文件生成任务上表现突出。
多代理编排让主导代理能将子任务委派给专业代理并行执行,共享文件系统,非常适合复杂调查。Harvey、Netflix和Spiral等团队已在用于法律工作、日志分析和写作代理。
链接:https://claude.com/blog/new-in-claude-managed-agents
Claude Blog: New in Claude Managed Agents: dreaming, outcomes, and multiagent orchestration.
Anthropic is launching "dreaming" as a research preview for Claude Managed Agents. This scheduled process reviews past sessions, extracts patterns, and refines memory so agents self-improve over time. Combined with persistent memory, it creates robust self-improving systems.
Outcomes let developers define success rubrics; a separate grader evaluates results independently, allowing agents to iterate until meeting criteria. This boosted task success by up to 10 points, with strong gains on file generation.
Multiagent orchestration enables a lead agent to delegate subtasks to specialist agents running in parallel on shared filesystems, ideal for complex investigations. Teams like Harvey, Netflix, and Spiral are already using these for legal work, log analysis, and writing agents.
Link: https://claude.com/blog/new-in-claude-managed-agents
查看原文 →