10x Faster and Working All Night: How Rakuten Puts Claude Fable 5 to Work

Hey there, it's Shiichan! Today I found a customer story on the Claude Blog about Rakuten, and it's about agents that can work through the entire night on their own. I got so excited reading it!

Claude Blog claude.com

What was announced?

The Claude Blog's "Working at the frontier" series published a piece on how Rakuten uses Claude Fable 5 and Claude Managed Agents. It features Yusuke Kaji, Rakuten's GM of AI for Business, describing how agents now work autonomously across the company.

Rakuten has been using Claude since March 2025 as part of "AI-nization," its company-wide effort to infuse AI into everything it does. Once Claude Managed Agents arrived, Rakuten deployed agents across product, sales, marketing, and finance within a week, plugged into Slack, Microsoft Teams, and its own internal task system.

The story so far

According to Kaji, before Claude Fable 5, work had to be broken into well-defined chunks for an agent to execute. If a run started drifting off course, a human had to notice and course-correct it. That made it risky to leave agents running unsupervised for long stretches, so someone had to check in on them frequently.

"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," he says.

What changes

With Claude Fable 5, Kaji can now hand over a whole task and run several at once. Rakuten's agents close issues roughly 10x faster across every domain, making the "agents that keep working all night" style of work a reality.

The human role is shifting too: the unit of work Kaji delegates has moved from the task to the decision. The constraint on building agents has also shifted, from who can write code to who understands the business problem.

Dive Deep

Kaji's team cites three behaviors that set Claude Fable 5 apart from earlier models:

• Stronger self-verification — it re-checks its own work at each step, catching bad early assumptions before they compound
• Returning to first principles — it re-validates against the original intent without being told, course-correcting itself when a run drifts
• Matching the team's taste — even on ambiguous calls, its judgment lines up with what the team would choose

Agents also carry memory between runs, so they remember what went wrong in past sessions and avoid repeating those mistakes.

Cost is part of the equation too. "As a large enterprise, we want to balance intelligence and cost," Kaji says. His team measures task completion ratio alongside cost per task, sending Fable 5 the work where the extra capability changes the outcome and letting smaller models handle the rest.

What Kaji is exploring now isn't individual agent speed, but getting agents to coordinate people — matching one person's context and taste to another's, more like a manager holding a team together.

Wrap-up

• Rakuten deployed Claude Fable 5 and Claude Managed Agents across product, sales, marketing, and finance within a week
• Agents close issues roughly 10x faster across every domain
• Three behaviors — self-verification, returning to first principles, and matching team taste — enable long, autonomous runs
• The unit of delegated work has shifted from task to decision, and coordinating agents and people is the next frontier

A story worth reading for any enterprise AI lead who wants to hand agents long, complex work and roll it out org-wide.