Claude Code team reveals 70-80% of daily work now delegated to Slack-native AI agent Claude Tag
In a video interview, the Claude Code team discussed how they use their own tools, Claude Code and Claude Tag, to accelerate product development. They reported that 70-80% of their daily work is now handled by Claude Tag, a Slack-native AI agent, a shift from a year ago when engineers read AI work logs line by line. The team emphasized a philosophy of setting goals and letting the model achieve them, with a 'scary forced放手' approach that decouples the UI from the model's internal reasoning.
Coverage timeline
量子位量子位
Claude Code团队讲究啊，这都往外说 衡宇 2026-09-17 19:56:13 来源： 量子位 工程师的核心永远是Problem Solving。 程浅 发自 凹非寺 量子位 | 公众号 QbitAI 话说。 Claude发布了一则Claude code团队如何使用Claude code持续优化Claude code的访谈视频。 访谈中，Claude员工们对最近这一年工作方式的变化大发感慨。 由此重点谈到了他们如何看待 “要在非常快的迭代节奏下做产品” 这件事。 △《How the Claude Code team uses Claude Code》访谈截图 （这画面温馨得像三个超级奶爸在讨论如何育儿。） 本文整理了这次访谈的核心内容，精华包括不限于： “你必须对自己正在构建的东西保持非常不执着的态度，因为它们很快就会消失。” “我们现在的工作层级，已经比Claude实际输出的token高出了一个抽象层。” “被迫放手让Claude自己发挥，因为信任比监督更重要。” ““你必须既站在技术前沿，又为今天使用模型的人提供价值。这是一门艺术，也是一门科学。” “软件工程是「变化的职业」。你解决的问题在变，但核心始终是problem solving。” 希望对同样处在 快技术迭代节奏下 的你有所启发。 团队现在是怎样工作的？ 团队表示，当前已经将70%-80%的日常工作交给Slack（美国飞书）原生的AI agent——Claude Tag完成。 而一年之前，Claude Code的工程师还在 逐行阅读 AI的工作记录。 包括每一次工具调用、每一个参数选择、每一步推理决策…… （这工作量。） 他们现在工作的核心理念是： 不再逐条审视tool call和模型决策，而是下达一个goal，让模型自行达成。 不知道这种理念是否与Claude Tag产品设计有关，或许是为了让它更像一个真实的Slack参与者，他们将用户界面和模型的思考记录彻底解耦了。 如此，AI的内部独白将被Slack界面隐藏起来，人类看到的每一条消息，只是Claude在调用发消息的工具，它真正的思考过程并不会实时展示。 团队认为这是一种 有点吓人的强制性放手 。 与此同时，他们也在激进地使用Claude Tag开发Claude Tag自身。 随之举了一个新内部工具的开发案例，很有意思： 第一步是问Cla
Related stories
Engineer describes company where all work is generated by Claude Code and unread
An anonymous engineer, quoted by Simon Willison, describes a large company where all work products—specs, code, tests, tickets, reports—are generated by Claude Code and nobody reads them. Employees from L1 to L7 work 12-13 hours daily just to submit AI output, while management wonders why progress is slow despite code not being a bottleneck.
1 reports · 1 sources
Anthropic's Claude Code drafts feedback reports for user review
Anthropic's Claude Code v2.1.238 or later can automatically draft feedback reports when tools fail, requests cannot be completed, or errors are identified. Drafts are saved locally at ~/.claude/feedback/drafts/ and are not sent to Anthropic until the user explicitly approves. Users can review, edit, or disable the feature via /config.
1 reports · 1 sources
Anthropic details Claude experiments in protein design, analytical chemistry; plans scientist access program
Anthropic published two experiments demonstrating how Claude can accelerate protein design and analytical chemistry, and announced plans for an access program for scientists. The results highlight Claude's potential to help life scientists increase the pace of their research.
2 reports · 2 sources
Anthropic releases findings from pilot letting external researchers study Claude usage; one study found users delegate high-stakes tasks
Anthropic released findings from a pilot that gave three external researchers access to aggregate, real-world Claude usage data. One study found that users delegate high-stakes tasks to the AI, highlighting emerging patterns in human-AI collaboration.
News updates every hour. Subscribe in your RSS reader to get new stories automatically.
Subscribe