Rakuten

Yusuke Kaji, General Manager AI at Rakuten, described the pattern in Anthropic's October 2025 launch of Agent Skills: multiple spreadsheets processed, critical anomalies detected, reports generated according to established procedures. Work that took a day now takes an hour.

Four months later, Anthropic's 2026 Agentic Coding Trends Report cites Rakuten again, this time for a different order of magnitude. A complete implementation ran for seven autonomous hours on a codebase of 12.5 million lines, at 99.9% accuracy. That figure anchors Trend 3 in the report, the shift from agents working over minutes to agents working over days.

The two appearances are worth reading together. The first is a workflow compression on ordinary office tasks, the kind of thing a skill packaged once and reused makes routine. The second is a long-horizon run on a codebase large enough that the accuracy number carries real weight: 0.1% error across 12.5 million lines is a different claim than 0.1% error on a small module, and the report does not break down what was measured.

Rakuten sits alongside Box, Notion, Canva, Fountain, TELUS and Zapier in the same reporting, an online commerce and technology company used as evidence of what «Claude Code» and agent tooling do at enterprise scale. What the sources do not say is how the seven-hour run was supervised, or how the 99.9% was verified.