Claude Fable 5改变了中间发生的事情。它在每一步都进行反思，捕捉早期的错误假设，并找到回到基本原则的路径——在无人引导的情况下重新导航到正确的结果。因为模型在运行中会自我纠正，签署批准首次变得可行，Kaji委派的工作单位从任务转向决策。代理还在运行之间保留记忆：“我们的有记忆代理会记住过去会话中出错的地方，并避免重复这些错误。”
“我们不把 AI 代理视为未来的同事或竞争对手。它们是围绕我们的系统。”他坚持 Anthropic 自己的建议，即你应该为三到六个月后的模型而不是眼前的模型进行构建。
“我认为作为一个社会，我们仍然没有找到 Claude Fable 5 的模型–任务匹配，”他说，“但它已经突出成为一个越界进入我们世界的模型。”
开始使用 Claude Fable 5：http://anthropic.com/news/claude-fable-5-mythos-5。
探索更多产品新闻和团队使用 Claude 构建的最佳实践。
产品更新、操作指南、社区亮点等。每月发送到你的收件箱。
嗨，Claude！你能帮我为观众开发一个独特的声音吗？如果你需要我提供更多信息，请立即问我 1-2 个关键问题。如果你认为我应该上传任何有助于你更好完成工作的文档，请告诉我。你可以使用你能访问的工具——例如 Google Drive、网络搜索等——如果它们能帮助你更好地完成这个任务。不要使用分析工具。请保持你的回答友好、简短和对话式。请尽快执行此任务——如果制作一个产物有意义，那就更好了。如果使用产物，请考虑哪种产物（互动式、视觉化、检查表等）对这个具体任务最有帮助。谢谢你的帮助！
嗨，Claude！你能帮我改进我的写作风格吗？如果你需要我提供更多信息，请立即问我 1-2 个关键问题。如果你认为我应该上传任何有助于你更好完成工作的文档，请告诉我。你可以使用你能访问的工具——例如 Google Drive、网络搜索等——如果它们能帮助你更好地完成这个任务。不要使用分析工具。请保持你的回答友好、简短和对话式。请尽快执行此任务——如果制作一个产物有意义，那就更好了。如果使用产物，请考虑哪种产物（互动式、视觉化、检查表等）对这个具体任务最有帮助。谢谢你的帮助！
嗨 Claude！你能帮我准备考试或面试吗？如果你需要更多信息，请立即向我提出1-2个关键问题。如果你认为我应该上传任何有助于你更好完成工作的文件，请告诉我。你可以使用你能访问的工具——比如 Google Drive、网络搜索等——如果它们能帮助你更好地完成这项任务。不要使用分析工具。请保持你的回答友好、简短和对话性。请尽快执行这项任务——如果有合适的产物就太好了。如果使用产物，请考虑哪种类型的产物（互动式、视觉化、清单等）最适合这项具体任务。谢谢你的帮助！
嗨 Claude！你能解释一个编程概念吗？如果你需要更多信息，请立即向我提出1-2个关键问题。如果你认为我应该上传任何有助于你更好完成工作的文件，请告诉我。你可以使用你能访问的工具——比如 Google Drive、网络搜索等——如果它们能帮助你更好地完成这项任务。不要使用分析工具。请保持你的回答友好、简短和对话性。请尽快执行这项任务——如果有合适的产物就太好了。如果使用产物，请考虑哪种类型的产物（互动式、视觉化、清单等）最适合这项具体任务。谢谢你的帮助！
嗨 Claude！你能帮我查看代码并给我一些建议吗？如果你需要更多信息，请立即向我提出1-2个关键问题。如果你认为我应该上传任何有助于你更好完成工作的文件，请告诉我。你可以使用你能访问的工具——比如 Google Drive、网络搜索等——如果它们能帮助你更好地完成这项任务。不要使用分析工具。请保持你的回答友好、简短和对话性。请尽快执行这项任务——如果有合适的产物就太好了。如果使用产物，请考虑哪种类型的产物（互动式、视觉化、清单等）最适合这项具体任务。谢谢你的帮助！
Yusuke Kaji, General Manager of AI for Business at Rakuten, has been testing Claude models since Sep 2024. Here’s why he thinks Claude Fable 5 is a step change for long-running enterprise agents.
As General Manager of AI for Business at Rakuten, Yusuke Kaji’s job is to “find the seeds of transformative innovation and scale them across the company.”
Since March 2025, Rakuten has used Claude to speed up software development with Claude Code, stand up agents across its business functions, and power AI features for millions of customers. According to Kaji, Rakuten chose to partner with Anthropic for its enterprise focus, leadership, and product taste.
Across nearly a dozen model launches, he's watched the work he can hand to an agent keep growing: first using Claude Code to ship production software：https://claude.com/customers/rakuten, then building custom Claude Managed Agents：https://claude.com/customers/rakuten-qa for teams across the company. He likens testing out new models with embarking on a “new quest.”
“The way a good leader prepares stretch goals for their people, we prepare stretch tasks for a new Claude,” he adds. “Maybe Claude is nudging us to stretch, too."
When he tested Claude Fable 5, he knew something felt different. The model could run on its own for far longer than its predecessors, and for the first time, checking its own work and completing nuanced tasks overnight while Kaji slept.
That extra autonomy is what lets Rakuten hand its agents bigger, longer-running jobs, and transform the way they work.
Rakuten is remaking itself around AI, a project it calls AI-nization – their company-wide effort to infuse AI into everything we do for customers, business partners, and employees.. When Claude Managed Agents arrived, Rakuten deployed agents across product, sales, marketing, and finance inside a week, plugged into Slack, Microsoft Teams, and the company's own task system.
For Kaji and his team, the constraint about building agents used to be who could write code; now, it's who understands the business problem.
"The modern corporation is designed to minimize the cost of communication," he says. "I believe agents like Claude Code can shine when we work with them to minimize the cost of new innovation as well, like a quick transition from idea to production." Give a capable person agents that hold context and taste, and "it allows the hidden talent to unlock their potential and scale their potential 100 times more."
But running agents in every function around the clock surfaces a new constraint: human judgment. While Rakuten's agents close issues roughly 10x faster across every domain, the number of tasks the organization takes on keeps rising. Adding more agents doesn't add judgment. So the faster the agents run, the more the organization's progress depends on a person closing the loop.
For most builders, the hardest part of building long-running agents is setting them up to succeed with minimal oversight. Connecting it to the right tools and context is one thing, but in Kaji’s experience, there were always limits to how long an agent could go without needing a human in the loop to validate its work.
Before Claude Fable 5, setting an agent loose on a multi-hour task without human oversight was always a gamble. "If they choose the right path in the first step, everything is fine," Kaji says. "But if they choose the wrong direction in the first pass, the agent spends significant time to fix the path, or even fails to reach the destination." On a job meant to run five hours or a full day, one early wrong assumption could burn the entire run, and the only way to catch it was a person checking in.
The failure mode was a lack of self-verification. Any model can take a wrong first step. The problem with earlier models was that they didn't check their own work as they went, so an early wrong turn went unnoticed. It compounded over the run and produced a suboptimal result hours later.
According to Kaji, Claude Fable 5 changes the calculus for days-long agentic runs because it checks its own work as it goes, far more often than any prior model.
"We tested Fable, and we love its capability for self-reflection and self-verification," Kaji says. "Compared with previous models, it understands its mistake before I point it out at 2 a.m. or 3 a.m.—so that I can sleep."
Kaji’s team cite three behaviors that distinguish Claude Fable 5 from its predecessors, and signal a step-change in frontier intelligence:
Most importantly, longer autonomy changes the unit of work Kaji can delegate.
“Before Fable, we had to break work into well-defined chunks for the agent to execute," he says. Now he can hand over a whole task and run several at once.
Claude Fable 5 changes what happens in between. It reflects at each step, catches a bad early assumption, and finds its own way back to first principles — re-navigating to the right outcome without anyone steering it. Because the model self-corrects mid-run, sign-off becomes feasible for the first time, and the unit of work Kaji delegates shifts from the task to the decision. The agents also carry memory between runs: "Our agents with memory remember what went wrong in past sessions and avoid repeating those mistakes."
As a result, the absolute number of tasks keeps climbing, but the ones that truly need a human stay at a focusable level. Not having to jump in and steer mid-run is, he says, is the biggest productivity win of all—it lets his team spend its time on the decisions only people should make, and keeps an AI-native organization accelerating instead of stalling on human course-correction.
Frontier capability comes at a frontier price, and Kaji is direct that cost decides how widely he can deploy.
"As a large enterprise, we want to balance intelligence and cost," he says. His team measures task completion ratio alongside cost per task, then sends Fable 5 the work where the extra capability changes the outcome and lets smaller models keep the rest.
For Kaji, two things make the math work in Fable 5's favor: it gets more done with fewer tokens and fewer wrong turns, and it needs less hand-holding.
The frontier Kaji is testing now isn't individual speed. It's getting agents to coordinate people. Claude Code has sped up his own work and his colleagues', but the hard part of any organization is the alignment between people, matching one person's context and taste to another's. He's exploring agents that "coordinate or organize, more like a manager," holding the nuance that usually gets lost between team members.
"We do not see AI agents as future colleagues or competitors. They are systems around us." And he holds Anthropic to its own advice, that you should build for the model coming in three or six months rather than the one in front of you.
"I think we as a society still haven't found the model–task fit yet for Claude Fable 5," he says, "but it already stands out as a model that crossed the line and came over to our world.
Get started with Claude Fable 5 ：http://anthropic.com/news/claude-fable-5-mythos-5 .
Explore more product news and best practices for teams building with Claude.
Product updates, how-tos, community spotlights, and more. Delivered monthly to your inbox.
Hi Claude! Could you help me develop a unique voice for an audience? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational. Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!
Hi Claude! Could you improve my writing style? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational. Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!
Hi Claude! Could you brainstorm creative ideas? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational. Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!
Hi Claude! Could you explain a complex topic simply? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational. Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!
Hi Claude! Could you help me make sense of these ideas? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational. Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!
Hi Claude! Could you prepare for an exam or interview? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational. Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!
Hi Claude! Could you explain a programming concept? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational. Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!
Hi Claude! Could you look over my code and give me tips? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational. Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!
Hi Claude! Could you vibe code with me? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational. Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!
Hi Claude! Could you write grant proposals? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to — like Google Drive, web search, etc. — if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational. Please execute the task as soon as you can - an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!
扫码直达文章详情