Rakuten(opens in a brand new window) is an international innovation corporate running throughout e-commerce, fintech, and cell communications, serving each customers and traders at huge scale. With 30,000 workers international, its engineering groups send throughout a big, advanced product ecosystem the place each velocity and reliability are crucial.
That’s why Yusuke Kaji, Common Supervisor of AI for Industry at Rakuten, has spent the previous 12 months pushing agentic workflows deeper into how groups plan, construct, and validate instrument. Codex—the coding agent from OpenAI—has turn out to be a core a part of Rakuten’s engineering stack, particularly the place the corporate wishes to transport sooner with out compromising safety.
Over the last 12 months, Rakuten engineers have used Codex throughout operations and instrument supply to compress incident reaction (together with a ~50% aid in imply time to restoration, or MTTR), reinforce CI/CD with computerized code assessment and vulnerability assessments, and toughen extra self sustaining building on advanced initiatives.
“We don’t simply care about producing code briefly. We care about transport safely. Pace with out protection isn’t good fortune.”
—Yusuke Kaji, Common Supervisor of AI for Industry
Inside of Rakuten’s engineering workforce, their AI schedule is crisp and deliberately operational. Kaji frames the paintings round 3 priorities that groups rally in the back of:
- Construct sooner (“Pace!! Pace!! Pace!!”): Groups use Codex in operational workflows, together with KQL-based tracking and prognosis, to boost up root-cause research and remediation, serving to compress MTTR via as much as 50%.
- Construct more secure (“Get issues accomplished”): Codex is invoked in CI/CD for code assessment and vulnerability assessments, making use of inner requirements routinely so groups can send briefly with guardrails.
- Perform smarter (“AI-nization”): Codex drives higher, ambiguous initiatives ahead from specification towards operating implementations, lowering dependence on perfectly-defined necessities, enabling extra self sustaining execution, and in the end compressing quarter-long efforts into weeks.
Codex maps without delay to each and every precedence as a unswerving agent in a broader toolkit, appearing up the place velocity, protection, and autonomy create compounding worth.
Pace at Rakuten comprises restoration time, no longer simply building speed.
Groups use KQL (Azure’s question gadget for logs and telemetry) to observe APIs and analyze alerts. Codex works along those workflows to lend a hand determine root reasons and counsel fixes, lowering the time between alert and backbone.
From a website online reliability engineering (SRE) standpoint, this shortens the trail from detection to remediation. As a substitute of manually sewing in combination queries, logs, and patches, engineers can center of attention on validating and deploying fixes.
Rakuten estimates this manner can scale back MTTR via roughly 50% when problems happen. Or extra merely put: Rakuten has used Codex to mend issues two times as rapid when one thing breaks.
As transport hurries up, assessment and deployment can turn out to be bottlenecks. Rakuten addresses this via integrating Codex without delay in its CI/CD pipeline.
Codex conducts code assessment and vulnerability assessments sooner than adjustments achieve manufacturing. Rakuten feeds inner coding rules and requirements into those workflows so opinions align with corporate expectancies.
“We offer our inner coding rules to Codex,” Kaji says. “The usage of the similar rules, it opinions whether or not the code aligns with our requirements.”
The outcome: protection assessments occur constantly and routinely, enabling groups to transport sooner with out decreasing requirements.
Rakuten’s 3rd precedence—AI-nization—specializes in autonomy. Codex is used no longer just for assessment and upkeep, but additionally for executing higher, ambiguous initiatives end-to-end. As a substitute of requiring completely explained specs, Codex can transfer ahead from partial necessities and convey usable artifacts.
“The newest Codex fashions can learn between the strains,” Kaji says. “Although the necessities aren’t completely explained, it understands what we’re seeking to construct.”
One instance: development a cell app model of an present web-based AI agent provider. Codex carried out all the specification, involving a complete stack implementation with a Python/FastAPI backend and a Swift/SwiftUI iOS app, together with all of the backend APIs, with out step by step human instruction. Codex reduce the improvement time for this mission from one quarter to weeks.
As Codex takes on extra code technology paintings, Rakuten is transferring the engineer’s position to writing clearer specs and verifying outputs towards measurable requirements.
“Our position isn’t to test each and every line of code anymore,” Kaji says. “Our position is to outline obviously what we wish and identify how to make sure it.”
Rakuten has supported this shift via hands-on workshops throughout engineering, product, and non-technical groups—contributing to Codex enjoying a central position in serving to groups send sooner, perform extra safely, and scale self sustaining building around the group.
Sign up for the brand new technology of labor
Greater than 1 million companies around the globe are attaining significant effects with OpenAI.
