📖 英文と日本語訳（一文ずつ）

🎧 通し読み（全文）

リスニング・シャドーイング用の全文です。

Anthropic announced on May 6, 2026, the release of "dreaming," a new feature for Claude Managed Agents, in research preview. Additionally, "outcomes," multi-agent orchestration, and webhooks have been made available to developers, aiming to realize agents capable of handling complex tasks with minimal instructions. Dreaming is a scheduled process that regularly reviews an agent's past sessions and memory store to extract patterns and organize memory. Users can choose to have memory updated automatically or to apply changes only after they have been reviewed by a human. Memory is reconfigured to maintain high information density by identifying recurring errors unnoticed by individual agents, converging workflows, and preferences shared across the team. Particularly effective for long-term tasks and multi-agent collaboration, the integration of memory functions for accumulating learned data and "dreaming" for organizing and sharing that information across sessions constitutes a coherent framework for self-improving agents. Outcomes is a feature that enables agents to self-correct their outputs toward specific criteria based on a rubric detailing success conditions. A key feature is that a separate "grader" evaluates artifacts in an independent context window, allowing for judgments that are not influenced by the agent’s reasoning. Should any issues arise, the grader will identify the areas requiring correction, and the agent will re-perform the task. It is reportedly effective for tasks where a "good state" can be defined—such as structural frameworks, presentation standards, and requirement sets—as well as for subjective quality assessments, including alignment with brand voice and design guidelines. Anthropic reported that internal testing showed task success rates improved by up to 10 percentage points compared to standard prompting loops, with the most significant improvements seen in more difficult problems. Success rates for file generation saw confirmed improvements of 8.4% for docx and 10.1% for pptx. Furthermore, users can now define outcomes and delegate execution, receiving notifications via Webhook upon completion. Multi-agent orchestration is a mechanism in which a lead agent divides tasks that are too heavy for a single agent and delegates them to specialized agents, each equipped with its own unique model, prompts, and tools. Specialized agents operate in parallel on a shared file system, reflecting their results into the lead agent's overall context. Due to event persistence, lead agents can also check the status of other agents mid-workflow. The Claude Console allows users to track which agents performed what actions and in what order, providing visibility into the task delegation and execution process. As a practical application, Harvey utilizes Managed Agents for legal workflows involving long-form drafting and document generation. The introduction of "dreaming" has enabled the system to remember file format workarounds and tool-specific patterns, resulting in an approximately sixfold improvement in completion rates. Netflix’s platform team has developed an agent that performs parallel analysis of logs from multiple sources across hundreds of builds to extract only the recurring issues found across numerous applications. Every’s Spiral utilizes a configuration where a lead agent powered by Haiku handles intake and simple follow-up questions, while draft creation is delegated to a sub-agent powered by Opus. This structure ensures that outcomes maintain a quality that aligns with editorial policies and the user’s writing style. Wisedocs has developed an outcome-based document quality check agent, reducing review times by 50%. Dreaming is available as a research preview, while Outcomes, multi-agent orchestration, and memory are offered in public beta as part of Managed Agents.

📝 学習のヒント

1. 1まず英文を読む — 知らない単語にあたりをつけてから音声へ。
2. 2一文ずつ確認 — 日本語訳と照合し、構文を理解する。
3. 3通し読み Normal で耳を作る — 内容を追いながらリピート。
4. 4Fast でシャドーイング — 口を慣らし、リスニング速度を上げる。
5. 5翌日に復習 — 1日空けて再聴すると長期記憶に定着しやすい。