Rakuten France, the French e-commerce marketplace subsidiary of Rakuten Group, runs 80+ Custom Agents inside Notion that automate more than 58,000 workflows across the company and cut backlog health issues by 50% on the product team.
"Our role isn't just to build automation solutions for the company—it's to act as a catalyst and spread best practices to every team," says Clément Caillol, CPO of Rakuten France. The result, he argues, is a "pragmatic productivity win: less manual coordination, more reliable execution, and better knowledge-sharing across teams."
Before the rollout, a four-person AI team had to build every automation the company needed, and previous no-code tools had failed because workflows broke whenever an employee changed roles. AI Lead Sidney Golstein anchored the rebuild in shikumika—roughly "systematization"—the principle that a workflow should produce the same result no matter who performs it.
Golstein began by authoring a set of Shared Operating Procedures (SOPs) that every new agent references, so common knowledge such as how to log a run, scrape data, or optimize instructions is captured once in Notion rather than re-invented per agent. On top of the SOPs, he built the Rakuten Agent Builder, a meta-agent that asks a few clarifying questions in plain language, drafts full agent instructions, automatically incorporates the SOPs, and generates a step-by-step setup guide. New agents typically go live in 10–20 minutes. Rollout took the form of in-room workshops where every team built their first agent live, with a Friendly Onboarder agent walking new users through what to expect.
Behind the agents sits an interconnected infrastructure. Every new Custom Agent is added to a central Agent Registry database so employees can discover existing workflows instead of duplicating effort. After every run, agents write a structured entry to a shared Agent Logs database capturing the trigger, actions taken, and any failures. A separate Custom Agent called the Delicious Improver reviews the logs to spot patterns, surface fixes, and propose updates to the SOPs—every suggestion then goes to the agent's owner for review, keeping humans in the loop. "We have a system where one Custom Agent creates agents; the agents that were created run and log; and another Custom Agent reviews the logs and improves the agent that just ran," Golstein explains. "That makes the system self-learning—a closed loop."
The product team offered the first production-scale test case. Three Custom Agents—Plan Snitch, Run Snitch, and Release Snitch—scan active tickets every night against a defined set of best practices and log their findings in a dedicated Notion database. A ticket stuck in review for more than three days or a release marked overdue is automatically flagged. Every morning, the Sprint Lebowski agent reads those findings and posts a tailored report to each team's Slack channel before tech leads, product managers, and release managers have logged on. Backlog health issues dropped measurably, with some teams cutting them by 50%; other teams, such as user care, asked to have Sprint Lebowski set up after seeing the daily Slack reports.
Today, 80+ agents run across Rakuten France, and most were not built by the AI team. Caillol describes the resulting shift as reflexivity—once equipped with a new capability, people start asking whether their own work is automatable and how it fits the larger system. "My job used to be building agents. Now it's making it easy for everyone else to build them," Golstein reflects, framing Custom Agents as the vehicle that moved Rakuten France beyond individual AI experiments and into a scalable way of working.