Rakuten engineering teams cut the average time to market for new features from 24 working days to 5, a 79% reduction, after redesigning their development workflows around Anthropic's Claude Code, according to Anthropic's customer story on the Japanese technology group, which runs more than 70 businesses across e-commerce, travel, fintech, digital content and communications.
"You can have five tasks running in parallel by delegating four to Claude Code while focusing on the remaining one," said Yusuke Kaji, General Manager of AI for Business at Rakuten.
With thousands of developers serving millions of customers, Rakuten measures AI tools on speed and ROI rather than raw code output. Its evaluation found that most AI coding tools needed constant human guidance and could not navigate complex multi-language codebases. The company's broader "AI-nization" strategy, which includes building its own agents and models, shaped how it assessed the tools.
The turning point was a test by machine learning engineer Kenta Naruse, who asked Claude Code to implement an activation vector extraction method in vLLM, an open-source library of 12.5 million lines in several languages. Claude Code completed the job in a single seven-hour autonomous run with occasional guidance, and the result matched the reference method with 99.9% numerical accuracy.
Rakuten then redesigned workflows rather than retrofitting AI into old ones. Engineers use Claude Code to write unit tests, mock APIs, build components, fix bugs and generate documentation, and new team members use it to learn unfamiliar codebases. Pull requests get AI-powered review, and developers run several Claude Code sessions in parallel to remove bottlenecks. Non-engineers also contribute through the terminal interface, with context and coding guidelines acting as guardrails. Rakuten tracks test coverage and an "AI-nization ratio" alongside delivery speed.
The company has since deployed Claude Managed Agents for product, sales, marketing and finance, plugged into Slack and Teams, with each specialist agent set up within a week. Naruse is building an "ambient agent" that splits a monorepo-wide update, normally more than a month of manual work, into 24 parallel Claude Code sessions, and Rakuten plans to extend Claude Code from quick tasks to multi-hour workflows across its developer organization.
Details
- City
- Tokyo
- Organization