Outcomes

Agent 很擅长交出「看起来做完了」的东西，也很擅长把记忆越记越乱。Anthropic 在 Claude Managed Agents 里给这两个问题各配了一个机制：Outcomes 在一次会话里请一位独立评审逐条验收，Dreaming 在两次会话之间把记忆重新整理一遍。本文讲清它们怎样工作、为什么要这样设计，以及怎样在自己的 Agent 里复刻。 版本范围：本文依据 2026-10-06 核对的官方文档。Outcomes 和记忆库处于公开 beta；Dreaming 是研究预览，要先申请才能用。两者都在 2026-05-06 的 Code with Claude 大会上发布。beta 阶段字段和限额都可能变，接入前以当时的文档为准。 最重要的一句提醒：Outcomes 的效果几乎完全取决于评分标准怎么写。 标准写得含糊，评审会什么都放行，循环一轮就结束，你付了钱却什么都没检查。 这篇文章写给谁 # 默认读者是这样的工程师：知道 LLM、上下文窗口和基础的工具调用，做过或配置过 Agent，平时用 Python、TypeScript、Go 之类的语言，但没用过 Claude Managed Agents。

Agents are good at handing in work that looks finished, and good at letting their memory drift into a mess. Anthropic gave Claude Managed Agents one mechanism for each problem: Outcomes brings in an independent grader to check the work inside a session, and Dreaming reorganizes memory between sessions. This article explains how they work, why they are designed this way, and how to rebuild them in your own Agent. Version scope: this article follows the official documentation as checked on 2026-10-06. Outcomes and memory stores are in public beta; Dreaming is a research preview that requires requesting access. Both were announced at Code with Claude on 2026-05-06. Fields and limits can change during beta, so check the current docs before integrating.