How Rakuten Uses Claude Code to Cut Feature Delivery from 24 to 5 Days

Rakuten’s scale creates problems that most engineering organizations never face. With over 70 businesses, thousands of developers, and millions of customers across e-commerce, travel, fintech, and digital content, the company needed AI capable of navigating genuinely complex, multi-language codebases—not tools that required constant hand-holding to produce incremental suggestions.

The company’s leadership had already invested in building its own large language models and AI agents, giving it an unusually clear-eyed view of what AI could and could not do. After evaluating existing AI coding tools and finding them unable to sustain independent work on enterprise-scale repositories, Rakuten tested Claude Code on a demanding benchmark: implement a specific activation vector extraction method in vLLM, a library containing 12.5 million lines of code across multiple programming languages.

Claude Code completed the entire implementation autonomously in seven hours— without a single line of code written by the engineer who initiated the task. The result achieved 99.9% numerical accuracy against the reference implementation. That outcome, combined with Anthropic’s alignment with Rakuten’s responsible AI values, moved leadership to commit to a company-wide rollout. When Anthropic released Claude Managed Agents, Rakuten deployed specialist agents across product, sales, marketing, and finance within a single week, with each agent integrated into Slack and Teams to handle long-running tasks including generating spreadsheets, presentations, and functional applications.

The impact on development velocity became measurable quickly. Average time to market for new features fell from 24 working days to 5—a 79% reduction. Engineers run multiple Claude Code sessions in parallel, delegating separate workstreams simultaneously. Non-engineers who previously had no role in technical projects can now use Claude Code via the terminal interface to contribute to coding efforts with appropriate guardrails. Individual engineers report qualitative changes too: test-driven development, once uncommon, became natural because Claude Code generates comprehensive tests instantly and then builds features that pass them.

Rakuten’s next step is an “ambient agent” that breaks complex tasks into 24 parallel Claude Code sessions, each handling a different aspect of updating the company’s massive monorepo—work that would otherwise take more than a month to complete manually. The company frames this not as automating existing tasks but as multiplying what each team can achieve, with technical barriers no longer the limiting factor for innovation.