Claude的新"Dreaming"功能旨在让AI代理从错误中学习

2026-05-07 18:59·56天前·Matthias Bastian

AI 摘要

Anthropic为Claude Managed Agents新增“Dreaming”功能，这是一个异步处理过程，能自动回顾历史代理会话、清理重复或过时的记忆条目，并提炼新见解。同时，Outcomes和Multiagent Orchestration两项功能现已进入公开测试阶段。这些更新旨在帮助AI代理在跨会话中从错误学习，实现持续改进，提升代理的自主学习和适应能力。

原文 · 未翻译

Claude's new "Dreaming" feature is designed to let AI agents learn from their mistakes

Anthropic is expanding its Claude Managed Agents platform, launched in April, with three new features for making AI agents more reliable and adaptive in real-world use.

A novel addition is "Dreaming," a process that reviews past agent sessions, spots patterns, and shares insights across sessions, such as recurring errors or proven workflows.

Technically, Dreaming runs as an asynchronous job: it reads an existing memory store and optionally up to 100 past sessions, cleans up duplicates and outdated entries, and builds a new, organized memory from them. The original memory stays intact. Claude Opus 4.7 and Claude Sonnet 4.6 are currently supported, and billing follows standard API token pricing.

Outcomes: Separate evaluator checks agent results against fixed criteria

Outcomes and Multiagent Orchestration are moving from research preview to public beta. With Outcomes, developers define a rubric: a document with specific success criteria like "The CSV file contains a price column with numerical values."

A separate evaluator (grader) then checks the agent's output against these criteria in its own context window, without being influenced by the agent's reasoning. If the result doesn't meet the specs, the grader identifies the gaps, and the agent revises its work, up to three times by default, with a maximum of 20 attempts.

Multiagent orchestration: Coordinator distributes work to specialized sub-agents

With multiagent orchestration, a lead agent (coordinator) manages the work of several specialized agents. Each agent runs in its own thread with an isolated context, its own model, system prompt, and dedicated tools, but they all share the same file system. The coordinator can hand out tasks in parallel, like delegating code review and test creation to different agents at the same time. The system supports up to 20 different agents and a maximum of 25 threads running simultaneously.

The Decoder：AI News（RSS）

59导出 Markdown